๐ฅ Top Story
Scope workloads and share resilience policies with AWS Resilience Hub
The next generation of AWS Resilience Hub gained three capabilities on September 18, 2026. The first is Amazon EKS label support for service input sources: within a namespace, selectors using In, NotIn, Exists and NotExists operators pick Kubernetes resources, and multiple selectors combine with AND logic. In the sample, app In checkout in the production namespace includes the checkout Deployment and ReplicaSet and excludes the inventory service, although shared infrastructure such as the cluster, node group and VPC still appears in the resource list. The second is generative AI-powered Dependency Insights, which analyzes up to 35 days of Route 53 Resolver DNS activity without an agent and flags cross-Region, new (first seen within seven days), third-party and uneven-usage patterns. For the example checkout service it produced five findings across 29 dependencies, including DynamoDB and PagerDuty reached in us-west-2, and insights can be regenerated only once every 24 hours. The third is resilience policy sharing through AWS Organizations, where a central team defines a policy once and shares it with member accounts; the example tier-1-critical policy sets 99.9% availability, multi-AZ RTO 5 minutes and RPO 1 minute, multi-Region RTO 15 minutes and RPO 5 minutes, and backups every 60 minutes. The article stresses that a policy-attached event shows adoption, not compliance, that connections made directly by IP are missed by DNS-based discovery, and that turning sharing off removes the policy from member services.
๐ก Why it matters: Pinning service boundaries to EKS labels and standardizing RTO/RPO through a shared organization policy can reduce drift between documented resilience requirements and what actually runs in the cluster.
๐ Read more ยท AWS Architecture
Kubernetes & Cloud Native
Why Cybersecurity Awareness Month still comes back to the basics
Sysdig
Sysdig's Crystal Morin argues that Cybersecurity Awareness Month still comes back to the basics, and that those basics are year-round work rather than a seasonal topic. The Verizon 2026 DBIR, covering more than 22,000 confirmed breaches, found vulnerability exploitation had become the top initial access vector at 31% of breaches, up 55% year over year, while third-party and supply chain involvement rose 60% to 48%. Google Cloud's Threat Horizons report found compromised identities in 83% of cloud and SaaS intrusions in H2 2025, and Sysdig's 2026 report says 97.2% of managed identities at cloud providers belong to machines. The post says 50% of AI tools run under the radar as shadow AI, and cites that 68% of organizations with an AI-related incident had no AI governance policy. Recommendations include MFA on all privileged accounts, short-lived credentials and least privilege, patching by exploitability rather than CVSS alone, continuous posture checks with minimal Kubernetes RBAC, runtime sensors with drift detection on every workload, centralized and long-retained logs, and signed artifacts in the supply chain. It concludes that these proven controls fail only when treated as an annual checklist instead of a daily discipline.
๐ก With machine identities dominating and most granted permissions going unused, tightening cluster RBAC and CI/CD credentials is likely the highest-return security work for cloud teams.
Docker Hub security advisory: personal access tokens transferred during account-to-organization conversions
Docker
Docker disclosed CVE-2026-105445, a flaw in Docker Hub's account-to-organization conversion that reassigned personal access tokens (PATs) to the new owner. Since August 26, 2024, conversion changed only the owner record, without rotating or revoking the token secret and without asking or notifying the new owner, so an attacker holding the secret could authenticate to the registry as the target user. Within the token's scope, that allowed pulling, pushing, and deleting images in the target's repositories and accessible organizations, but not web UI access, password, email, or 2FA changes, or account takeover. Docker received the report on September 30 and deployed a fix in under four hours, emailed exposed users on October 2, and force-deactivated the transferred PATs on October 8. It reviewed about 304,000 pushes made with transferred tokens and found no evidence of exploitation. The fix removed PAT transfer from conversion entirely, and Docker plans to require consent and notification for owner assignment and to keep an auditable history of token ownership changes. Users should review their PATs, revoke unknown ones, and use organization access tokens for automation.
๐ก Tokens tied to personal accounts can become an exposure path through administrative side effects like ownership transfers, so CI and automation should use organization tokens and registry auth logs should be reviewed regularly.
From zero to agent builder: Running kagent on your laptop
CNCF
In a CNCF blog post, Antonio Berben (Solo.io) and Felipe Vicens (Telefonica) build a builder agent on a laptop using kagent, a CNCF Sandbox project that runs AI agents as Kubernetes custom resources. The setup installs kagent v0.10.2 with Helm on a kind cluster named kagent-builder and uses Ollama with the qwen3-coder:30b-a3b-q4_K_M model, about 18 GB; the smaller qwen2.5 7b and 14b models were unreliable at tool calls. The agent-builder is allowed only three tools, and k8s_apply_manifest is gated by requireApproval, so a human must approve each Agent manifest in the UI before it is applied. The same flow can be scripted over A2A with the JSON-RPC message/send method, where an input-required status means approval is pending. Guardrails are layered: the prompt, a tool allowlist, human approval, and RBAC, where the tool server gets a read-only role plus one extra rule that lets it write agents.kagent.dev and nothing else, and it cannot read Secrets. The default unsecure auth mode and the bundled PostgreSQL's built-in password make this unsuitable for a shared cluster, and the author proposes agentgateway as a governance layer, with a follow-up post planned.
๐ก When agents are just YAML resources, creating them is cheap, so without up-front governance over who requests, approves, and pays for them a cluster will accumulate unmanaged agents.
AI & ML
Run fast, private agents on-device with LFM2.5-24B-A2B
Hugging Face
Liquid AI released LFM2.5-24B-A2B on October 13, 2026, an open-weight reasoning model with 24B total and 2B active parameters per token, aimed at fast, private agents on local hardware. It has 40 layers, 64 experts per MoE block with top-4 routing, and a hybrid of gated short convolutions and grouped-query attention, with the context window raised from 32K to 131K tokens and the vocabulary from 65K to 128K. Pre-training grew from about 17T to about 33T tokens, and post-training has five stages ending in agentic reinforcement learning inside harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode. Across 16 benchmarks against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning, it leads on tool use and instruction following, with scores such as AutomationBench 27.69, BFCLv4 65.01, ToolSandbox 80.71 and IFEval 96.75. Larger models still lead on knowledge and coding, for example GPQA Diamond 74.42 and SWE-Bench Verified 51.00 for this model, and the article suggests pairing it with retrieval or web search for knowledge-heavy work. Reported speed is 184 tokens per second on an Apple M5 Max and 114 tokens per second on a Ryzen AI Max+ 395, and over 21K output tokens per second at high concurrency on a single NVIDIA B300. Day-one support covers llama.cpp, MLX, vLLM, SGLang and ONNX.
๐ก If a 2B-active-parameter MoE can run tool-calling agents on laptop-class hardware, local automation that never sends data off the machine becomes worth evaluating for operational workloads.
Gemini 4 Argon: Built to solve your hardest problems
Google AI
Google announced Gemini 4 Argon on October 13, 2026, the first release in the Gemini 4 family, aimed at deep multi-step work in coding, finance, legal, cybersecurity and science. The API is available in AI Studio and the Gemini platform with a 1M unified input-output token limit, at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens 95% off. After November 30, 2026 the public version rises to $4 per million input tokens and $20 per million output tokens. On benchmarks it ranked first on the Vals Index covering finance, coding, legal and tax work, scored 87% on ERQA, and cut attack success on the Gray Swan indirect prompt injection benchmark to 0.7% across 15 adaptive attempts. Engineering examples include a rise of more than 35% in agentic code submissions in Antigravity with nearly a third fewer rollbacks, and months of Chromium refactoring work, separating core webpage handling from a roughly 6-million-line legacy folder, reduced to weeks. Throughput for Gemma 4 E2B rose from about 37,900 to over 108,000 tokens per second, and science examples include an optimized hardware kernel for the AlphaQubit decoder and rare-disease genome analysis. On safety, Google applied its Frontier Safety Framework, external bio red teaming could unlock less than 1% of the bio sub-tasks tested, and a model card was published.
๐ก Because the introductory price doubles after November 30, teams planning agentic workloads should weigh token costs before and after the increase, along with prompt injection resistance, when deciding when to adopt.
Higgsfield takes AI video from 1 minute to over 10
OpenAI
OpenAI's Higgsfield customer story is titled as taking AI video from 1 minute to over 10, but the page itself returned HTTP 403, so the details below come from press coverage and search extracts of it. Per Blockchain.News and TokenPost, Higgsfield says it extended AI-generated video from roughly one minute to more than 10 minutes using OpenAI models including GPT-6 Astra. According to an extract of the OpenAI page found via search, co-founder and CEO Alex Mashrabov says videos topped out around one minute before GPT-6 Astra, and that its agents can now call external tools for 3D modeling work, which he says helps with consistency across longer videos. Higgsfield also cites GPT-Image-2.5 for more precise visuals and GPT-Live as a step toward voice filmmaking, which the coverage describes as a future direction rather than a shipped product. The material reviewed contains no benchmarks, pricing or rollout details, and it is unclear whether the longer-video capability is generally available.
๐ก If long-form video is now an agent workflow with tool calls rather than a single model call, operators will need to plan for long-running job state, retries and per-job cost visibility.
Cloud Updates
Under the hood, off the hot path: How D1 scales to millions of databases
Cloudflare
Cloudflare explains how the D1 control plane manages millions of databases, reporting that as of October 8, 2026 live databases grew 15x over one year and 65x over two years, and daily active databases grew 18x over the past year. The central principle is to keep centralized, single-region components off hot paths such as queries, splitting work into a data plane at the edge, a control plane that may depend on single-region components, and an asynchronous plane that can be eventually consistent. Each D1 database is backed by one SQLite Durable Object, routing metadata lives in Workers KV, and central metadata lives in Postgres reached through Hyperdrive. Every Durable Object sends a heartbeat to the control plane every 30 seconds to sync database state and account limits, clean up orphaned objects and correct Workers KV entries, which makes the system self-correcting. Storage limits are 5 GB on the Free plan and 1 TB on Workers Paid; a job aggregates a change log every second into account totals, and ClickHouse is the billing source of truth. Accounts that exceed daily row read or write limits are found by a component called Limiter, which writes an expiring key to Quicksilver that Durable Objects check at the edge to reject queries. The post sums up the split as Postgres for transactional metadata, ClickHouse for analytics, and Workers KV for globally replicated routing data.
๐ก Solving the hot path with edge caches and self-healing heartbeats while pushing heavy work such as limit enforcement onto asynchronous paths is a practical reference model for keeping control plane failures from spreading to the data plane on multi-tenant platforms.
Cross-cloud conversational analytics made simple with open standards
Google Cloud
A Google Cloud blog post shows how its borderless Lakehouse, built on the open Apache Iceberg REST catalog specification, lets BigQuery query data in place across Databricks, Snowflake and AWS Glue without copying it. It argues that the traditional approach of extraction pipelines into a central warehouse creates synchronization lag, recurring storage costs, variable network egress fees and pipelines that break on upstream schema changes. The Lakehouse runtime catalog syncs schemas and snapshot pointers on a user-defined schedule, then at query time reads Parquet files directly from remote object storage such as Amazon S3 using temporary credentials issued by the remote platform, with all processing running in BigQuery or Managed Service for Apache Spark. The walkthrough registers Unity Catalog, Snowflake and Glue catalogs with gcloud alpha biglake iceberg catalogs create, joins a native fact_sales table with dimension tables from all three clouds in one query, and saves it as a view for natural-language conversational analytics. For production, the refresh interval has a minimum of 300 seconds and an overrunning sync causes BigQuery to skip the next scheduled run, and cross-cloud data blocks are cached in the catalog region, so teams should confirm this fits their data residency policies. If the organization enforces the constraints/gcp.restrictNonCmekServices policy, BigQuery bypasses the cache and reads remote storage on every query, which can be checked by comparing cacheBytesRead with objectStorageBytesRead in the job statistics.
๐ก Moving from ETL copies to Iceberg catalog federation lowers pipeline upkeep, but cross-cloud egress per query and the interaction between caching and CMEK policy need to be measured up front for cost and compliance.
Gemini 4 Argon available on Google Cloud, built to help solve your businessโ hardest problems
Google Cloud
Google Cloud announced that its new frontier model, Gemini 4 Argon, is available to everyone on Google Cloud. The API model ID is gemini-4-argon, and it is also reachable through Gemini for business, Google AI Studio, and Google Antigravity. Introductory pricing through November 30, 2026 is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 afterward, with cached input tokens priced 95% lower. ServiceNow reported roughly 20% better task success than Gemini 3.8 Flash in early access on NowAI-Bench, and Google says Argon leads the Vals Index and Harvey's Legal Agent Benchmark. Inside Google, teams using Argon in Antigravity raised agentic code submissions by more than 35% and cut rollbacks by nearly a third, while the Chrome team finished an architectural refactor across hundreds of files in Chromium's 40 million-line codebase in weeks. On security, Argon ties for first on CWE-Bench v1 with 68% and cuts attack success to 0.7% across 15 adaptive attempts on the Gray Swan indirect prompt injection benchmark; offensive tasks are blocked, but vetted defenders in the Fairwind Program get a version without cyber guardrails.
๐ก Because list prices double after November 30, agent workloads should not lock in a cost model at the introductory rate and should re-estimate spend including cache hit rates.
Rewinding the exponential: How we keep hardware usage down
Cloudflare
Cloudflare's Data team runs the pipeline that moves data from its global network to central data centers, and it describes five years of changes that offset about 2.5 years of hardware expansion against exponential traffic and linear hardware growth. Delivery ran at roughly 2 billion events per second in May 2026; splitting a monolithic schema into product-specific ones in 2022 cut backbone bandwidth by 35-40%, and switching from LZ4 to zstd in 2024 cut compressed throughput by nearly 40%. In 2026, load balancing replaced double-writing every non-regional event to two regions with a shared ingest pool and late replication downstream of Buftee, for a net saving of 36 GB/s. Double sampling added to Logfwdr in summer 2026 draws a sample sized at the square root of each group's count, so a key with 100x more data stores about 4x the rows instead of about 10x, saving about 25 GB/s on the top three datasets. In-pipeline aggregation this fall cut the HTTP rollup dataset from over 9 GB/s to under 2 GB/s, about 80%. Ready-Analytics receivers saw about a 50% drop in CPU and memory use, regaining about 750 CPUs and 3.5 TiB of memory. Across spring to fall 2026 about 70 GB/s was removed from the backbone, throughput now trails the old growth projection by roughly eight months, and projected 2031 data volume was cut in half.
๐ก Cutting data volume as early in the path as possible, through compression, replication timing, sampling, and pre-aggregation, also shrinks downstream CPU and memory needs rather than just bandwidth.
Stopping business email compromise with AI-driven email security: How VIPRE protects modern enterprises on AWS
AWS Architecture
VIPRE Security Group built VIPRE Integrated Email Security, a cloud-native email security solution on AWS, to stop business email compromise (BEC). The post cites VIPRE data that 49 percent of spam emails are BEC scams, and argues that Secure Email Gateways relying on MX record redirection can miss internal traffic, lateral movement, and thread hijacks. The product connects to Microsoft 365 through secure APIs with no MX record changes. Each email is stored in S3, a job is queued in SQS, and stateless ECS/Fargate services run static, behavioral, semantic, and contextual analysis, with custom BERT models on ECS with NVIDIA GPUs for semantic detection. Events flow through MSK into RDS as the system of record, and change data capture streams them to OpenSearch with results scoped by tenant ID. In a 24-hour sample of EU and North America production, 517,726 messages were scanned and 5,508 BEC attempts were detected, about 1 percent. P95 latency is 1.4 to 15 seconds without dynamic sandboxing and 39 to 78 seconds with full sandbox analysis, at over half a million messages per day.
๐ก API-based email security avoids MX changes, but sandbox analysis stretches latency to tens of seconds, so post-delivery clawback needs to be designed alongside inline blocking.
Bulk Executor for Amazon DynamoDB, Part 2: Lessons we learned the hard way
AWS Architecture
This Part 2 post covers the design tradeoffs and surprises behind Bulk Executor, an open-source tool that runs commands against many items in a DynamoDB table. After weighing SQS with Lambda, Fargate, EMR, and Ray, the team picked AWS Glue for serverless parallelism and its built-in DynamoDB connector, which made find nearly free through Spark DataFrames, though the connector cannot distinguish a string set from a list of strings. Ctrl+C originally stopped only the local script while the Glue job ran to timeout, so an interrupt handler now tells Glue to stop; CloudWatch Live Tail samples beyond 500 events per second, so large output goes to S3 and the terminal shows the first 10 results plus a link. Point-in-time recovery is required for mutating commands and revert-export can undo writes, and because the connector cannot delete, delete uses direct batch_writer calls, whose throttling was solved with a distributed rate limiter across workers plus XMaxReadRate and XMaxWriteRate options. The new scancount command scans 100 segments in parallel with Select COUNT and ran consistently faster than the connector-based count. The Glue 4 to Glue 5 migration was finished in one interactive session using AI-generated stress tests, the move to Glue 6 was made to capture about 30% lower DPU-hour costs, and an AI linter built from markdown rule files checks consistency such as IAM permission descriptions. The project lives at awslabs/amazon-dynamodb-tools under tools/bulk_executor.
๐ก For bulk tooling against live tables, safety rails such as clean interruption, upfront cost estimates, rate limiting, and a rollback path matter more than the feature list.
One Agent Development Kit, five languages
Google Cloud
Karl Weinmeister of Google Cloud explains that the Agent Development Kit (ADK) supports five languages, Python, Go, TypeScript, Java, and Kotlin, each built for its own runtime and type system. Every language includes an MCP client and supports Agent2Agent (A2A), so agents in different languages or frameworks can interoperate over HTTP. Getting started uses uvx google-agents-cli setup and agents-cli scaffold create, which generate a project with A2A endpoints and deployment config for Cloud Run, Gemini Enterprise Agent Runtime, or GKE. The Go version fits a roughly 2 MiB distroless static base image, goroutines start with a 2 KB stack so one instance can multiplex thousands of streaming requests, and it is the first language after Python to ship the ADK 2.0 Graph Workflow runtime. TypeScript uses Zod schemas and can run in a browser tab using WebMCP and Chrome's on-device Prompt API. Java offers bridges for Spring AI and LangChain4j, while Kotlin uses KSP to generate tool schemas at build time, survives Android R8 shrinking, and can run the full agent loop on-device with LiteRT-LM and ML Kit without a network connection.
๐ก Since MCP and A2A are the shared layer, platform teams need not standardize on one agent runtime and can pick per component, for example Go for high-concurrency services on Cloud Run.
Whatโs old is new: Lightwell discovers novel zero-day vulnerability in core Java framework
Red Hat
A Red Hat blog post argues against the assumption that code untouched in production for ten years must be safe, since decades-old foundational libraries such as the Eclipse Implementation of Jakarta XML Binding (JAXB) quietly power enterprise applications. Red Hat's Lightwell team discovered a novel vulnerability in JAXB, identified as CVE-2026-92877. The post says advanced AI security tooling does not ignore these deep-stack layers and finds zero-day flaws in structural parsing logic that standard static analysis tools miss entirely. It lists the risks of leaving such a flaw unaddressed: parsing bugs chained with other vulnerabilities to escalate privileges or execute code, malformed payloads causing unhandled exceptions or resource exhaustion, and supply-chain amplification because JAXB is buried in thousands of enterprise packages. Lightwell contributed the fix upstream to the eclipse-ee4j community and calls upstream acceptance of the patch a major milestone. For older releases that upstream communities no longer maintain, it says Lightwell backports security patches so organizations are protected without disruptive, unplanned upgrades.
๐ก Scanning images for known CVEs will not catch novel flaws in deep transitive dependencies like JAXB, so teams should track dependencies with SBOMs and weigh backport support when choosing vendors.
How AI is defining the โnewโ architect
Red Hat
Red Hat's Idan Henik argues that as AI and autonomous agents automate code and infrastructure generation, the architect's value shifts from owning the technical blueprint to governing the business and operational consequences of AI-generated systems. He calls the new role the Cognitive Architect, combining human judgment, enterprise context, platform thinking and governance. Among seven reasons, he says prompt-driven vibe coding creates risk because AI does not understand legacy dependencies, compliance duties or cost limits, and that production AI brings GPU scarcity, token economics, inference latency and multi-tenant isolation concerns. Governance is framed as a four-stage A-framework of Awareness, Appreciation, Acceptance and Accountability, supported by Policy as Code, OpenShift AI and human-in-the-loop decision frameworks. He warns that automating away junior debugging work risks organizational cognitive atrophy, and proposes Code-as-a-Textbook, built on adversarial code reviews between junior engineers and senior architects. The piece promotes Red Hat Consulting services and offers no customer case studies or external statistics.
๐ก As AI-generated infrastructure gets faster, embedding automated guardrails such as Policy as Code into delivery pipelines becomes the platform team's central job.
Filter, reshape, and redact logs before export with Transformers
Cloudflare
Cloudflare Logpush now supports Transformers, which let teams filter, redact, and reshape logs with SQL before export. Each Transformer is a SQL query for one Logpush dataset: a WHERE clause drops records, the SELECT output becomes the exported record, and fields can be added, renamed, removed, nested, or masked. Input and output are both NDJSON, a Transformer attaches to a compatible NDJSON job, every update creates a new version, and they are managed in Transformer Studio or through the API. They run inside Bufstream, Logpush's Go delivery service, where Apache DataFusion in Rust executes the SQL on Apache Arrow data; Cloudflare rejected a WebAssembly-in-Workers design because the network call added timeouts and retries. In a synthetic benchmark a prepared Transformer processed a 5,000-row http_requests batch in about 3 ms on one goroutine, roughly 1.6 million rows per second, while Bufstream delivers more than 110 million Logpush records per second overall. Transformers fail closed: on error the offset does not advance, the untransformed batch is not sent, and Bufstream retries, and jobs that keep failing are disabled automatically. They are available on Free, Pro, Business, and Enterprise plans.
๐ก Masking personal data and dropping health-check noise at the source before logs reach a SIEM or storage can cut both ingestion cost and compliance exposure.
Developer-first innovation: Reducing technical debt with Red Hat Enterprise Linux
Red Hat
In this post, Red Hat product manager Louis Imershein starts from the tension between developers wanting freedom to pick tools and leadership needing security and stability, and recommends standardizing on a few enterprise-grade operating systems. He argues that when every team runs a different distribution, developers spend sprint hours on infrastructure, and maintaining base Linux layers creates a DIY maintenance tax that turns senior engineers into unpaid OS maintainers. RHEL is pitched as providing backported security fixes that do not break downstream dependencies, and the Red Hat Developer Subscription for Teams is said to charge nothing for development, CI/CD test pipelines or staging instances, billing only production workloads and a few internal tools. Image mode for RHEL, built on bootc, defines the whole OS in a single Containerfile in Git, and the post says its read-only immutable images and atomic updates with automatic rollback reduce configuration drift and mean time to recovery. On supply chain security it presents Red Hat Hardened Images and Red Hat Trusted Libraries, and for InfoSec gates it cites pre-certified CIS Level 2 profiles, SELinux and FIPS 140-2/3 validation. For local workstations it covers the Red Hat build of Podman Desktop, kaiden for secure agentic sandboxes in Red Hat Desktop, and Red Hat Lightspeed included with the RHEL subscription.
๐ก Treating the OS as a container image built in CI brings node patching and rollback into the normal delivery pipeline, though the benefit has to be weighed against vendor lock-in and subscription terms.
DevOps & Infrastructure
Honeycomb Is Hiring a New VP of Engineering
Honeycomb
Honeycomb announced in a blog post by Charity Majors that it is hiring a new VP of Engineering, and the post says the company has never before posted a job opening on its blog. Emily Nakashima, who has led engineering for more than nine years, is moving into an operational role, and the new hire will partner with SVP Product Graham Siener to lead R&D. Honeycomb is remote-first with about 300 people, 85 of them in engineering, and the role reports to the CEO as a full member of the executive team. Responsibilities include shaping strategy across product, technology and market, building alignment with product and the CEO, helping the organization make sense of change, leading the existing managers, directors and ICs, and keeping technical credibility by having shipped software hands-on with AI in the past 12 months. The three core needs are a partner to product, enough risk tolerance for pre-product-market-fit offerings, and an AI-forward, hands-on background. Strong preferences are having led an organization of hundreds or more, having owned all of engineering, and having served on an executive team, and the post encourages candidates who match only some criteria to apply through a Greenhouse job board link.
๐ก An observability vendor making hands-on AI experience and an agentic roadmap core requirements for its engineering leader is a signal to weigh a tool's direction on AI workloads when choosing observability platforms.
GitHubโs model for secret detection: Finding more secrets with less noise
GitHub
GitHub introduced a context-aware classifier for generic secret detection that raised owner-weighted acceptance by 11.5% and cut median post-push scan execution time by 86% compared with its previous LLM pipeline. Serving cost is more than 98% lower than remote LLM calls, and offline evaluation showed 13% higher precision and 19% higher recall. The design has two stages: high-recall regex patterns generate candidates, then a fine-tuned ModernBERT model scores each one using the candidate value, file path, secret type and surrounding text. The hard problem was telling apart strings that look identical but differ by context, such as a password next to a production database URL versus the same value in a test fixture, or an API credential versus a checksum, build ID or asset hash. Training data came from product feedback with human validation, human-annotated and weakly labeled regex candidates from repositories, and synthetic examples aimed at documentation secrets, test fixtures, public keys and checksums. The model is exported to ONNX with 8-bit quantization, and the lower scan time made it practical to extend generic secret detection to push protection. Future work includes a feedback flywheel that turns production false positives and negatives into training data, and models that perform detection and extraction in a single step.
๐ก A small fine-tuned classifier sitting behind cheap regex candidate generation shows that secret-scanning noise can be cut without paying per-call LLM costs, which matters for how much alert triage load your pipelines and on-call engineers carry.
Anthropicโs red team ran AI on 15 cheap TV boxes. Then the math got unsettling.
The New Stack
For roughly $825, Anthropic's Frontier Red Team built a cluster of 15 off-brand Android TV boxes that runs an AI model at 0.39 tokens per second, too slow for a chatbot but fine for software with nobody waiting on it. The experiment anchors research by William Zhou and Daniel Freeman suggesting small open-weight models may be only months away from copying themselves across cheap, internet-connected hardware without a datacenter. They have not built a worm that spreads across the public internet, and the tests ran in controlled environments with shortcuts that made replication easier. The researchers split replication into four steps, finding a target, gaining access, getting inference running and moving on, and see inference on unfamiliar hardware as the biggest hurdle. Qwen3.5-9B deployed a working Llama3.1-8B inference server on eight of nine hardware types, including Jetson Nanos, Macs and RTX 4090 machines, and in an eight-node Google Cloud Android emulator setup GLM-5.2 reached all eight nodes in four of five trials. When agents could read notes left by earlier attempts, which the researchers call Lamarckian inheritance, Qwen3.5-9B success rose from 45% in the first generation to between 61% and 84% in later ones. Defenders have levers too: moving multi-gigabyte models creates traffic ISPs can watch, and during the Mirai cleanup one ISP's quarantine cleared 92% of infections within 14 days.
๐ก If self-installing an inference server on unfamiliar devices is the real bottleneck, then egress traffic monitoring and network observability around IoT and edge fleets become the first line of defense.
Metaโs Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI
Meta Engineering
Meta presented the 2026 generation of GEM, its ads recommendation foundation model, moving from separate per-domain models for Facebook, Instagram and Reels to a single central model trained on data and signals from all of its ads platforms. Reported results are an 18.9% conversion improvement and a 12% click-through rate improvement, with additional conversion uplift for targeted users over the past year of 5.8% on Facebook, 4.1% on Instagram and 1.6% off-site. Model complexity grew 2.5x over earlier GEM versions, and the Kunlun architecture delivers 2x the performance of the previous Interformer architecture at the same computation scale. Block Attention and Generalized Dot-Product Attention are designed around GPU kernels, with GDPA giving up to a 2x forward speedup, while Component Skip Connection cuts FLOPS by 43% with negligible quality loss. For transferring knowledge to downstream models, LoopFM doubles the transfer gained from the previous distillation approach and Solaris, inspired by speculative decoding, precomputes user-ad embeddings in asynchronous background jobs. Total training FLOPs grew 4x while end-to-end training efficiency doubled to 20 to 25% model FLOPS utilization across thousands of GPUs, using MXFP8 low-precision training and topology-aware 5D parallelism. Future directions include learning from text, images and video, LLM-scale long context, and adding reasoning to post-training.
๐ก Because Meta attributes its 2x training efficiency to co-optimizing kernels, precision, parallelism, networking and memory together, it is a case showing that GPU cluster cost savings come from whole-stack tuning rather than any single layer.
Gemini 4 Argon debuts with a unified 1M input/output token limit
The New Stack
On Tuesday Google made Gemini 4 Argon, its new flagship, available to developers on the paid tier of the Gemini API through Google AI Studio, after announcing it on September 30 with initial access limited to selected cyber defenders in its Fairwind Program and other trusted testers. The introductory price through November 30 is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 afterward to match Claude Opus 5.5, with cached input tokens 95 percent off during the introductory period. During that window it costs a fifth of Claude Fable 5.1 and GPT-6 Astra, which both list at $10 and $50. The key change is a unified 1 million token budget for input and output combined, where the September 30 announcement had described a 1 million token output cap, so a request loading 800,000 tokens of code leaves only 200,000 for the response. Claude Opus 5.5 by comparison pairs a 1 million token context window with a separate 128,000 token output cap. Developers call it as gemini-4-argon through the Interactions API, while Antigravity access is limited to Google AI Ultra subscribers and eligible Gemini Enterprise plans. Google reports 87 percent on ERQA and Artificial Analysis scored it 53 on its Intelligence Index, while Bloomberg reported that some Google employees saw it struggle with certain coding tasks, which Google called inaccurate.
๐ก Because input and output now draw from one shared 1M-token budget, long-context agent workloads need prompt-size guardrails, and cost models should already assume the post-November 30 price.
The agent did the work. The audit log couldnโt say who asked.
The New Stack
More than 20 percent of the MCP-related access policies that Shawn Petty's team reviewed across customer and prospect environments were broken or missing, according to an account published in The New Stack. In most cases the MCP server ran on someone's personal token rather than a service account. A personal token puts an engineer's name in the audit trail for everything the agent did, and a shared service account puts the account's name there instead, so in both cases the trail cannot say who asked the agent or who approved the access. Access that goes through a request, approval and grant in the identity provider is provisioned and recorded, whereas access an agent picks up along the way, such as a role inherited from the launching service or a developer's lent token, is emergent and leaves no record in the identity provider. This page promotes a live demo on Tuesday, October 27 at 12 p.m. Eastern with CloudBolt's Shawn Petty (Chief Customer Officer) and Mike Bombard (Director of Global Pre-Sales Engineering), hosted by The New Stack's Viktor Farcic. The session covers tying agent activity to an existing identity provider, which infrastructure actions an agent may take without human sign-off, what a usable audit trail contains, and a live trace of one agent's credentials back to the responsible person.
๐ก Running agent MCP servers on personal tokens or shared service accounts erases the requester from your audit log, so agents need their own identities and a request-approve-grant path through the identity provider.
๋์์ด๋์ ์์ด๋์ด๊ฐ ๊ธ์ต์ํ์ด ๋๊ธฐ๊น์ง
ํ ์ค
Jang Seoyoung, a Product Designer at Toss Bank, shares how a designer's idea became a financial product: the 'Grow It 31-Day Savings' account, a revamp of the existing 'Grow It Savings' product. It began in the design chapter's OOTB (Out of the Box) time, prompted by the problem that once auto-transfer was set up the character grew on its own and users had no reason to open the app. Game worlds such as repairing a spaceship were dropped because they had no natural link to saving, and two revisit mechanics, a different story each day and multiple endings, were attached to the growing concept instead. With a PO, the team compared reward structures at the same cost and settled on up to 10% annual interest for saving 5,000 won a day, and user tests showed people with little savings experience were relatively positive about raising a character. To win over the team, the designer and a data analyst found that users holding savings visit the app more often and that many users only use a checking account, positioning the product as an entry-level savings product. Reported results are that half of subscribers save at least three weeks a month, one in five saves every one of the 31 days, and many who finish come back to raise a new character.
๐ก An experience improvement ships only when it is tied to the team's goals through cost structure and data, and the same applies when pitching internal platform or developer tooling work.
์๋น์ค ๊ฐ๋ฐ์ ์ง์คํ ์ ์๋ AI ํ๋ซํผ ๋ง๋ค๊ธฐ
ํ ์ค
Toss's AI Platform team describes how it moved the repeated work of model integration, permission checks, usage accounting and failure triage out of individual service teams and into a shared platform. The pieces are GenAI Portal for access requests and price/performance comparison, GenAI Gateway as the LLM gateway handling auth, permissions, usage limits, guardrails and provider adapters, MCP Gateway that checks internal tool execution rights on every call, Toss Dynamo for serving models on in-house GPUs, and Langfuse for tracing. Permissions, cost and usage are grouped per project, prompts are screened for attacks and personal data before any external model call, and personal data is masked in logs. Budget is managed as cumulative cost while request admission is limited separately by tokens per minute (TPM), and traffic goes to pre-purchased capacity (PTU) first, falling back to pay-as-you-go (PAYG) on the same model when it runs short. Data is classified into four levels (L1 to L4) and operations into three SLA tiers (High, Standard, Flex), and a data level alone never grants permission because purpose, environment and model are reviewed together. As a financial service, the team also checks each use against the scope allowed under Korea's innovative financial service designation. The post gives no performance or cost-saving figures and says evaluation criteria for response quality and cost will come in a follow-up.
๐ก Separating budget (cumulative cost) from throughput (TPM) at the LLM gateway, with PTU-first routing that spills to PAYG, is a reusable pattern for controlling both cost and availability in an internal AI platform.
How Datadog defines service criticality at scale
Datadog
Datadog describes a system that classifies internal services by criticality, defined strictly as how much customers are affected when a service is unavailable, slow or incorrect. Each service gets a level of high, low or unknown plus a total confidence score showing how much evidence backs it, with a medium level planned as more signals arrive. The signals are incidents extracted by an LLM from SEV-1/SEV-2 postmortems, APM and RUM error traces, production traffic, and monitors tagged to the service, where traffic can only lower a level and monitors that page an on-call rotation raise it. The incident pipeline processed several hundred postmortems from an archive of thousands in under a day for about $32, with extracted names validated against Software Catalog and a second model reviewing the first. Two independent agreeing signals count as stronger evidence, an opposing signal lowers confidence without changing the level, and service owners can override a level with a stated reason. Consumers use the result to deploy non-critical services faster in tiers across data centers and to auto-merge dependency updates on low-criticality services. Stated limits include reliable services that never fail being invisible to the incident signal, low-traffic services that sit on critical user journeys, and dependency mapping still being under development.
๐ก Deriving service criticality from incidents, traces and monitors and exposing it with a confidence score lets teams safely vary rollout speed and auto-merge scope per service.
Make every AI credit count: See, cap, and reduce agent spend
GitLab
GitLab's Allie Holland argues that agent costs grow with every task handed over, so platform teams need to see, cap and reduce spend. The first lever is matching models to tasks: GitLab 19.4 makes GitLab-managed open-weight models GLM 5.3, Kimi K3 and MiniMax M3 available in Duo Agent Platform, and per GitLab's internal benchmarks they matched task completion at about one-fifth the cost of Claude Opus 5, one-third of Claude Fable 5.1 and one-eighth of GPT-5.5 respectively. The second lever is GitLab Orbit, a continuously updated graph linking code, work items, merge requests, pipelines, deployments and security findings, which in GitLab's testing cut retries on the same task by up to 45x and used 4.5x fewer tokens to get the right context. The third lever is measurement and limits: Impact Analytics breaks spend down by team, task and model, the GitLab Credits dashboard traces usage to individual users and sessions, and caps can be set per subscription, group or user with alerts at 50%, 80% and 100% and an automatic pause at the cap. One example is a flat 1,000 credits per user per month with a 500-credit override for an experimenting engineer. Open-weight models and credit controls are generally available, Orbit is a free beta on GitLab.com with GA planned for the following month, and Impact Analytics is in early access.
๐ก Agent spend balloons from retries and context hunting more than from unit price, so set caps and alerts first, then pin cheaper models to routine workloads and compare spend by team.
GitLab's agent orchestration eliminates the wait between development stages
GitLab
GitLab announced agent orchestration for Duo Agent Platform at GitLab Transcend, in a post by Elisabeth Burrows. The premise is that AI sped up writing code but the bottleneck moved downstream to verify, review and ship, where reviewer sign-offs, pipeline alerts, scan triage and step-by-step prompting still wait on humans. Goal-driven flows take an objective and a definition of done, then run through review, tests, security scans and approvals, with a separate model checking the work against the goal at each step and the option to stop or redirect at any time. They start with /goal in Duo CLI, duo run --goal in headless mode, or Agentic Chat. GitLab for Slack lets you mention @GitLab to start a flow such as fixing a failing pipeline and streams progress back into the thread. Custom Flows define multi-step agent sequences once and trigger them from a mention, an assignment or a pipeline, for example a self-healing pipeline that analyzes a failed test, commits a fix and notifies the team with human checkpoints at sensitive steps. All flows run under GitLab's existing identity, policy and evidence chain, and goal-driven flows need Premium or Ultimate, GitLab 19.3 or later and Duo CLI 9.17.0 or later.
๐ก Once agents self-heal inside CI/CD, the permission scope of automated fix commits and the human approval checkpoints need to be defined up front as pipeline policy.
The โalienโ economics of AI monetization
Stripe
Stripe's James Brown recaps the second Monetize conference, which drew more than 300 leaders, where Metronome CEO Scott Woody called leading AI companies' operating style alien compared with traditional SaaS: faster pricing iteration, market share over ARR, and several go-to-market strategies at once. Two-thirds of Forbes AI 50 companies use some form of usage-based pricing, yet instead of replacing subscriptions, usage is increasingly packaged inside subscriptions through credits. Fin (Intercom) moved from per-resolved-conversation pricing to a growing set of billable outcomes drawn from a credit pool, and Replit abandoned 25-cent agent checkpoints once costs diverged, only to face unpredictable spend and then add a rate-limited Free Mode and a simple subscription. Predictability is treated as a product feature, with Snowflake's forecasting tools and calculators and what Stripe's Sam Lee calls a usage control pane with alerts and spending limits becoming an emerging standard. Clay's finance head Karan Parekh says Clay treats tokens as pass-through costs and measures AI spend by operational outcomes rather than token-level ROI. Kyle Poyar analyzed 7,600 AI-generated responses across six engines and found that when pricing was not accessible, LLMs pulled from sources like Reddit, producing a 24x range of $100 to $2,400 per user per month.
๐ก Teams running AI services should design credit pools and usage alerts and limits as part of billing itself, and publish pricing in a form crawlers and agents can read.
OCP Summit 2026: Powering AI with Open Innovations in Scale-Up, Scale-Out, and Scale-Across Networks
Meta Engineering
At the 2026 OCP Global Summit, Meta presented open AI networking infrastructure organized into scale-out, scale-across and scale-up. It introduced IcePack, its first liquid-cooled network rack, and IceCube, 102.4 Tbps switch blades with 128x800G capacity built on either Broadcom Tomahawk6C or NVIDIA Spectrum-6 ASICs, both running Meta's open-source FBOSS with OCP SAI at twice the bandwidth of its 51.2T platforms. The scale-out Flattened Planar Fabric attaches each GPU through eight 100G links to eight separate switches, collapsing a 3-tier Clos into 2 tiers, with the NIC spraying packets across all planes and NVIDIA's Spectrum-X Multiplane used in its development. For scale-across, a Backend Aggregation (BAG) super-spine joins fabrics across buildings, campuses and regions into one cluster, with inter-region capacity reaching tens of petabits per second. For scale-up, Tahan, an Ethernet switch for the MTIA 300 built on Tomahawk 5 and delivering 6.4 Tbps per accelerator, is shipping today, while eight of the previewed 216x400G blades connect 72 accelerators per rack and a future generation of the blade is planned to adopt OCP ESUN. ESUN, which Meta founded within OCP, is advancing open specifications alongside AMD, Arista, ARM, Broadcom, Cisco, HPE, Marvell, Microsoft and NVIDIA, and the Meta, Microsoft and Alibaba led OSPA submitted its v1.0 switch platform specification to OCP. Meta also published FBOSS self-onboarding documentation and the open-source RDMA transport MetaRoCE.
๐ก As switch OS, platform layer and transport all move to open standards, AI cluster networks become easier to operate and observe like software without vendor lock-in.
GitHub Copilot app for Beginners: How to teach Copilot about your codebase
GitHub
This post in GitHub's 'GitHub Copilot app for Beginners' series, written by Christopher Harrison, explains four ways to store project knowledge in a repository so Copilot picks it up automatically. Custom instructions put repository-wide rules in .github/copilot-instructions.md and path-specific rules in .github/instructions, with an applyTo field that says when each file applies. Agent skills are folders of instructions and optional scripts under .github/skills; the example is quality-checks/SKILL.md, which could describe the test and lint scripts and how to troubleshoot failures. Custom agents are profiles in .github/agents that define purpose, instructions, tools, and context, and they suit recurring work that spans files, such as an accessibility review. MCP servers connect Copilot to tools and data outside the repository, and the Playwright MCP server lets Copilot open the app in a browser to verify UI behavior. The author suggests starting with one convention in .github/copilot-instructions.md and points to Awesome Copilot, a community collection of instructions, skills, and agents. The article lists no limitations or benchmark numbers.
๐ก Keeping instructions, skills, and agent profiles as versioned files lets a team review AI coding conventions through the normal PR flow, though any MCP server you connect needs its own access-scope review.
Shipping Media Over QUIC at Meta Scale for Instagram Live Ingest
Meta Engineering
Meta says Instagram Live now uses Media over QUIC (MoQ), an open IETF transport, for live video ingest on iOS and Android. It replaces RUSH, Meta's in-house QUIC-based protocol that had taken over from RTMP for nearly a decade. Clients use moxygen, Meta's open-source MoQ library, together with mvfst, its QUIC stack, and a single config parameter selects MoQ or RUSH per broadcast. A temporary edge layer translates MoQ frames into the RUSH format so the backend stays unchanged, and server-side stream URL generation assigns broadcasts to the experiment regardless of app version. The hurdles were a MoQ draft that kept changing, UDP/443 being blocked or rate-limited on some enterprise and carrier networks, and QUIC encryption hiding framing, which led Meta to build MoQ-aware tracing at the edge. After months of experiments, MoQ matched RUSH on the quality and reliability metrics used for the rollout, and the article reports no latency or stall improvements. Future work includes end-to-end playback, a standards-based TCP fallback through QMUX, and a bidirectional transport for real-time AI.
๐ก The migration worked because the experiment was server-controlled and reversible with one config switch, and any QUIC adoption needs a fallback for UDP-blocked networks and a way to observe encrypted traffic before rollout.
CPU ๋ผ์๊น์ง ์จ๋จน๊ณ , Milvus ์ฌ์ด์ง ๊ธฐ์ค์ ๋ง๋ค๊ธฐ๊น์ง
ํ ์ค
Toss Securities' Data Infra team needed to size servers for on-prem Milvus vector search, so it benchmarked three servers (dual-socket Intel Xeon Silver 4516Y+, 96 logical cores and 512 GB each) with VectorDBBench LAION-100M, which has 100 million 768-dimension vectors. The target was top-10 search with recall@10 of at least 0.95 and at least 300 QPS, and only HNSW_SQ8 (ef48, recall 0.954, 316 QPS) met it, using about 114 GB per replica versus about 335 GB for plain HNSW. Raising concurrency from 10 to 300 lifted QPS only about 31% and IPC was low at about 0.37, and PMU counters showed roughly 49% of DRAM reads going to the remote socket on all three nodes. Using the kubelet CPU Manager static, Memory Manager Static, and Topology Manager single-numa-node policies to keep QueryNode CPU and memory on one NUMA node cut remote DRAM to 0.8% and raised IPC to 0.55. At the same recall, QPS rose 41% from 405 to 572 and search p99 fell about 48%, from 3.94 s to 2.04 s. The team built a sizing method that takes the larger of the server count needed for the QPS target and for memory, informed by their finding that doubling the dimension cuts throughput to about a third. The authors caution that LAION data differs from real service conditions, so measurements must be repeated on production data before adoption.
๐ก Because CPU utilization includes memory stalls, a throughput plateau calls for checking IPC and remote DRAM ratio, and NUMA alignment alone can unlock large gains on the same hardware for memory-bound workloads.
GitLab Critical Patch Release: 19.4.2, 19.3.4, 19.2.8
GitLab
GitLab released critical CE/EE patch versions 19.4.2, 19.3.5, and 19.2.8 on October 12, 2026, and strongly recommends that self-managed installations upgrade to the patch for their branch as soon as possible. The release fixes 24 CVEs: 1 Critical, 5 High, 14 Medium, and 4 Low. The Critical one is CVE-2026-97229 (CVSS 10.0), an integer overflow in the CI/CD regex compiler that allows unauthenticated code execution on the server, affecting versions from 19.2. The High issues include XSS in syntax highlighting (CVE-2026-92883), improper authorization in downstream pipeline creation (CVE-2026-16301), and LLM prompt injection in the GitLab Duo Agent Platform (CVE-2026-96893). Single-node instances will have downtime because migrations must finish before GitLab starts, while multi-node instances can use zero-downtime upgrade procedures. GitLab.com is already patched and GitLab Dedicated needs no action; 19.3.4 broke merge request pages when the Ona integration was enabled, and 19.3.5 fixes that. Issue details become public 90 days after release.
๐ก With unauthenticated remote code execution reachable through the CI/CD path, self-hosted GitLab needs immediate patching rather than the regular upgrade cycle, and single-node setups should schedule a downtime window up front.
if(kakao)26, ์จ๋ผ์ธ ๋ผ์ด๋ธ๋ก ํจ๊ปํด์!
์นด์นด์ค
This post from Kakao's DevRel Zoo says if(kakao)26, held at Kakao AI Campus on October 13 and 14, can be watched online as a live stream. All presentation sessions, including the keynote, are streamed live and anyone can watch online. Day 1 is organized as Keynote and Developers in the AI Era, Trust and Safety, AI Direction, Win-Win and Growth, and Model and Agent, and at the time of the post these live sessions are marked as ended. Day 2 runs through an opening speech, Governance, Stability and Infrastructure, AI-DLC, and Model and Agent. The event shares the Kakao team's technology and experience around the themes Intelligence in Every Connection on day one and The Moment Technology Becomes Trust on day two. Readers are told to check the official if(kakao)26 website for session times in advance, and recordings of the sessions are planned to be provided after the event for those who cannot watch live.
๐ก The Day 2 tracks on governance, stability and infrastructure, and AI-DLC line up with the operational questions cloud and DevOps teams hit when adopting AI, so they are worth prioritizing once recordings are available.
This digest was collected from RSS feeds and summarized by AI (Claude). See the original links for full details.