2026-11-11
Top stories
- Microsoft fixes 1,093 flaws on November Patch Tuesday, including exploited Outlook spoofing bug CVE-2026-124194 Microsoft's November update fixes about 1,100 flaws, and CISA lists an Outlook spoofing bug among those already exploited in attacks.
- SGLang code execution flaw let an attacker breach NEAR AI inference servers, including TEE-protected requests An attacker used an SGLang remote code execution bug to run code on a NEAR AI inference server and could see requests inside its TEE.
- Node.js 26.13.1, 24.22.3 and 22.23.5 fix 16 CVEs in HTTP/2, TLS and the permission model Node.js security releases for the 26, 24 and 22 LTS lines fix 16 CVEs, seven rated high, in HTTP/2 stream handling and TLS.
- AWS releases AWS Emulator, a free open-source Rust emulator of AWS APIs for local testing AWS Emulator serves AWS APIs on localhost to the CLI, SDKs and coding agents, with per-service fidelity scores checked against real AWS.
Security
- Many Electron apps auto-update without signature checks while still trusting compromised eMudhra CAs A survey of 36 Electron apps finds many auto-update without signature checks while still trusting compromised eMudhra intermediate CAs.
- cap-primitives 3.4.6 and 4.0.3 fix a critical sandbox escape through paths and symlinks ending in a slash A trailing slash let symlinks inside a cap-std sandbox open files outside it on macOS, Windows and Linux without openat2.
- eMudhra revokes eight more intermediate CAs whose RSA keys passed through a Python script's memory Six TLS and two code-signing CAs fell after the company's incident report said it could not prove their private keys stayed controlled.
- Grafana 13.3.1 and backports to 13.2, 13.1, 13.0 and 12.4 fix at least six CVEs Grafana shipped security releases on 2026-11-10 across five release lines, each listing six to eight CVE fixes without details.
- Apache httpd 2.4.71 fixes critical mod_http2 race CVE-2026-119618 that can lead to remote code execution A synchronization bug in mod_http2's assign_headers corrupts the heap in httpd 2.4.55 through 2.4.70, and 2.4.71 also fixes a mod_md crash.
Outages
AI
- Microsoft Research's Lazarus backdoors LLM agents by editing fewer than ten dormant neurons, with no training Microsoft researchers show that editing under ten rarely firing neurons backdoors an LLM agent without training while it still passes evals.
- DeepMind releases AlphaGenome 2, a cell-type conditioned DNA model for variant effect prediction DeepMind's AlphaGenome 2 predicts how DNA variants act in specific cell types and adapts to new single-cell data without retraining.
- Google launches the Gemini Decisions API, scoring up to 1,000 closed-set questions per request at $0.10 per million tokens Gemini Decision 1 returns calibrated probabilities for yes/no, choice and rubric questions in 150 to 250 ms, and output tokens are not billed.
Agentic coding
- Sentry's CLI now launches its MCP server with shared credentials, combining MCP tools with filesystem access Sentry's CLI now launches its MCP server over stdio with the CLI's own credentials, so agents get native MCP tools and filesystem access.
- Agents wrote and tested all of AWS Emulator's code, behind a merge gate with adversarial security agents AWS says agents built its emulator from 400+ API models, filing gaps found against real AWS and gating changes on security agents.
ML research
- Stochastic rounding closes BF16's slower convergence against FP16 in RL post-training Round-to-nearest leaves small RLVR updates stuck in BF16 weights, and stochastic rounding matches FP16 speed with lower KL drift.
- Shared parameters can suppress or amplify reward hacking in RLVR, depending on how many prompts accept a hack In an exact toy setting, a leak most prompts accept grows under RL, while a rare one shrinks below the base model's rate.
- Pyrite benchmark shows LLM agents beat expert chemists at ranking protein-ligand poses A pose-ranking benchmark finds expert chemists near chance on 1,500 ligand pose pairs, while LLM agents that write feature code reach 74%.
- Theory paper finds REINFORCE-based post-training can collapse coverage and shows KL reference resets prevent it A theory paper shows REINFORCE-based PPO and GRPO can drive optimal-response probability toward zero, and KL reference resets mitigate it.
Developer tools
- Val Town is winding down, going read-only on 2026-12-10 and shutting down on 2027-01-10 Val Town stops signups now, goes read-only on 2026-12-10, shuts all services on 2027-01-10 and plans to open-source the code on 2027-02-01.
- Atuin 18.24 adds Threads, a synced, encrypted history of agent sessions that you can resume in another agent Atuin Threads syncs Claude Code, Codex, Pi and OpenCode sessions with end-to-end encryption, and resumes or forks them in a different harness.
- Zellij 0.46 can also run as a standalone desktop terminal on Linux, macOS and Windows The zellij terminal command runs Zellij as its own desktop terminal app, sharing sessions with the multiplexer in other terminals.
Languages and runtimes
Apple platforms
Infrastructure
Engineering posts
Talks
- Agent Substrate reaches 1.0, and Anthropic adds a bridge to run Claude Managed Agents on it A KubeCon keynote demo packed about 65 suspended and resumable agent sandboxes per node on a 24-node GKE cluster.
- Meta runs one AI on-call teammate per rotation, with a three-level autonomy ladder and team memory Meta's agent triages alerts, drafts fixes behind human review and closes low-criticality work alone, while high-criticality stays human.
- Duolingo's agent pipeline merged 700 test PRs adding 225,000 lines of iOS tests, with a human approving each one Temporal workflows pick untested Swift files, Claude Code writes tests, and an LLM judge rejects PRs that touch production code.
- Apple Container adds container cluster, a kubeadm Kubernetes cluster of lightweight VMs on a Mac Apple showed container cluster create at KubeCon, running each Kubernetes node as its own VM with a choice of CNI and shared images.
Hacker News
- Hacker News debates VulnCheck's count of 442 Anthropic-credited CVEs and 217 fixes in Project Glasswing VulnCheck counts 442 CVEs credited to Anthropic in seven months, with 2 exploited and 217 of 37,966 ledger findings marked fixed.
- Hacker News argues over a list of open source projects that ban AI-generated contributions Maintainers and users split over whether bans protect review time or drive away contributors, in an 87-comment thread on osssaysno.com.
- Hacker News swaps datacenter war stories after a false smoke alarm nearly set off sprinklers in Jane Street's server room A smoke alarm shut down the cooling and pressurized the sprinklers, so an engineer pulled every power cord as the room reached 143 F.
- Guido van Rossum announces that Tim Peters, author of the Zen of Python and Timsort, died on 2026-07-15 Python's core developers remember Tim Peters, who designed Timsort, wrote PEP 20 and tuned pymalloc, dict and the cyclic garbage collector.