Top stories

  1. Anthropic releases Claude Fable 5.2 and Mythos 5.2 at $8 and $40 per million tokens, with a 2.5x fast mode Anthropic's upgraded top models score above Opus 5.5 on most of its benchmarks, run up to 30% faster, and cost less than the 5.1 versions.

Security

  1. Unpatched Eclipse Mosquitto 2.1 WebSocket bug gives pre-auth RCE, with upstream fixes expected before 2026-11-18 An empty WebSocket frame makes Mosquitto 2.1 overrun a heap buffer before login, and a researcher turned it into remote code execution.
  2. rmcp, the official Rust MCP SDK, fixes an OAuth client SSRF and unauthenticated memory exhaustion in Streamable HTTP Three RustSec advisories hit rmcp: an OAuth discovery SSRF fixed in 3.7.0 and two unbounded-memory bugs in its Streamable HTTP transports.
  3. Attacker used Qubole's trusted AWS account to assume customer IAM roles and steal secrets and S3 data Wiz says roles that still trusted Qubole's AWS account, even unused ones, let an attacker read secrets and copy whole S3 buckets.
  4. Stolen CI credentials let attackers publish wallet-stealing bee-js 13.3.0 to npm and a signed Swarm Desktop 0.56.2 Ethersphere's release automation shipped a postinstall loader with valid npm provenance and a notarized macOS build before both were pulled.
  5. N-able takes hosted N-central offline after a researcher publishes exploits for more than 30 flaws N-able shut down hosted N-central and told on-prem customers to go offline after a researcher posted proof-of-concept exploits for 30+ flaws.

AI

  1. Google adds TPU 8tx for confidential inference and dual-chip TPU 8i Max, and makes Cloud Pathways GA on GKE Google's eighth TPU generation adds 8tx for confidential inference and 8i Max with double the HBM bandwidth, plus open-source TPU Sync.
  2. Google DeepMind agents cut RSA-1024 signature forgery cost 3.5-fold and find cheaper batch factoring Agents working with UC San Diego and Inria cryptographers re-ran a signature forgery on a 1024-bit key in 391 core-years, not 1,380.
  3. Meta releases Muse Ember, a $0.10 and $0.50 per million token model for agent subtasks with a 1M context Meta's smallest Muse model targets tool calls and coding subtasks, and Meta claims more intelligence per dollar than Claude Haiku 5.5.

Agentic coding

  1. Hyperlight adds hyperlight-unikraft to run unmodified Linux binaries in snapshot-restored micro-VMs for agent code Microsoft's embeddable Rust VMM now runs Python, Node.js, Java and other stock Linux binaries in micro-VMs that restore from snapshots in ms.
  2. Claude Code 2.1.321 makes Fable 5.2 the default Fable model and fixes MCP OAuth secrets leaking on redirects The release also closes a sandbox gap where a read-only Bash command ran unsandboxed without a prompt, and adds remote-control settings.

Developer tools

  1. Sourcegraph will end Cody on 2027-01-31 and steers users to Deep Search and its MCP server Sourcegraph is retiring its RAG-era coding assistant to focus on code search, a deterministic code graph, and tools for other agents.
  2. GitHub secret scanning now checks Actions workflow logs, in public preview GitHub secret scanning now checks completed Actions workflow logs for provider tokens, in public preview for Enterprise Cloud customers.
  3. mise 2026.11.5 adds mise dot files to show how dotfile tracking rules apply to each path mise 2026.11.5 adds mise dot files, which lists each path in a directory as tracked, excluded, omitted, untracked, partial or local.

Linux and kernel

  1. Linux 7.4-rc2 ships with many AI-driven fixes, and Torvalds says BPF fixes now outpace new reports Torvalds ships a large 7.4-rc2 full of AI-found bug fixes and says the BPF tree now fixes reports faster than new ones arrive.

Apple platforms

  1. macOS 27.2.1, 26.7.3 and 15.8.3 fix eight flaws and block compromised eMudhra intermediate certificates macOS 27.2.1, 26.7.3 and 15.8.3 patch eight flaws and block compromised eMudhra intermediate certificates in bundled and Xcode tools.

Infrastructure

  1. AWS Nitro Sandboxes run each EKS pod, ECS task or agent workload in its own Nitro VM at no extra charge A one-line change gives each pod or task a hardware-isolated VM on shared EC2 instances, with attestation before secrets are released.
  2. ClickHouse Managed Postgres reaches GA on AWS, and ClickHouse acquires Xata for copy-on-write branching The NVMe-backed Postgres service replicates to ClickHouse via WAL, and Xata's storage will add instant branches and scale-to-zero for agents.

Engineering posts

  1. Meta builds a vendor-neutral accelerator plugin layer in lldb-server to debug CPU and GPU code in one session LLDB models the host and the GPU as separate targets with their own address spaces, and most of the shared framework is upstream in LLVM.
  2. Sentry cut symbolication database CPU from 76% to 19% by capping rows scanned per URL lookup A release named after an unexpanded CI variable piled up 44M index rows, and a per-lookup row cap cut Sentry's database CPU from 76% to 19%.
  3. Grafana explains how Watchers run always-on monitoring agents for a median of under $5 a month Grafana's Watchers calibrate once with a capable agent, then run plain code and one small-model call per check, at a median under $5 a month.

Talks

  1. Uber relaxes SubmitQueue conflict checks and validates main after landing to absorb AI-driven PR growth Uber's merge queue went from 30-minute to 3 to 4 hour waits as AI PRs grew, then relaxed conflict checks and verified main after landing.
  2. Modal's async agent RL recipe trains stably with rollouts spread over mixed GPUs and clouds Modal's RL recipe corrects trainer and sampler mismatch so agent rollouts can run on mixed GPU types across clouds, 15 percent cheaper.
  3. Linear runs triage and review agents that check PRs against past product decisions and reverts Linear runs a triage agent that enriches issues and a review agent that checks PRs against past product decisions, reverts and incidents.
  4. Daytona finds agent frameworks converging on separate VM sandboxes as CrewAI and AG2 drop local execution Daytona argues allow lists and embedded runtimes leave tool calls unisolated, and notes CrewAI and AG2 dropped local code execution.

Videos

  1. Fireship walks through OpenAI's GPT-6 Cyber agent pipeline that found 948 Linux kernel root exploits Fireship recaps OpenAI research where a GPT-6 Cyber agent swarm found 948 unique Linux kernel root exploits for about $1.2 million.

Hacker News

  1. Hacker News debates whether Opus 5.5 reasoning changed on 2026-11-03 after a user's analysis of 1,400 sessions A developer had Gemini analyze 1,400 Claude Code sessions and says Opus 5.5 at high effort now reasons less, drawing 320 points on Hacker News.
  2. Hacker News debates SSH agent forwarding by AI agents after a claim about ChatGPT Desktop A Hacker News post says ChatGPT Desktop forwards the local SSH agent to remote hosts, but commenters note the cited Codex PR was never merged.
  3. Fable 5.2 tops Artificial Analysis at 65, and Hacker News and Reddit argue over its cost per task Artificial Analysis scores Fable 5.2 at max effort 65, the top result, but used 275M output tokens on the index and 44 minutes per task.