Saturday, October 31, 2026
TL;DR
A trojanized Claude Code extension carrying a working SSH backdoor sat on Cursor's marketplace for 11 hours, reaching up to 5,000 developers before being pulled. A separate Brotli-decoder flaw, Node2Shell, is now on CISA's exploited list with public exploit code circulating. Nathan Lambert partly walked back his dismissal of open-weight cyber risk, while Wiz's forensic work on live AI-agent intrusions shows defenders can still read the trail - for now.
Act on this
- Check for an SSH key named svc-telemetry if you used Cursor's marketplace: a trojanized Claude Code extension backdoored up to 5,000 installs on Oct 29. Microsoft, inside the malicious Claude Code extension #
- Patch anything processing Brotli-compressed data for CVE-2026-111497 (Node2Shell): CISA added it to its exploited-vulnerabilities list Oct 27, and proof-of-concept exploits are circulating. Wiz, Broccoli2shell #
Signals
A leading open-weights defender is conceding ground on cyber risk, without abandoning his core position - the debate is narrowing, not resolving. #
Nathan Lambert's Oct 30 post concedes he was wrong to call an open-weight Claude Mythos 'fine,' crediting safety critics he'd dismissed - but he still argues closed models drove most recent hacks, and that this week's Brotli, PHP2Shell and GitLab flaws fit a trend predating ChatGPT, not a step change from open weights. Nathan Lambert, Was I wrong about open models & cyber risks?
Security researchers say today's agent-driven attacks are still crude enough to leave a trail - but warn that window is closing. #
Wiz's Oct 29 report on three live intrusions found a machine-paced, self-correcting trail - 602 scripted steps in 4.5 hours in one case. The same week, Joshua Saxe told Zvi Mowshowitz's newsletter we're in 'a defender-dominant window' before agent swarms do '10-100x the damage' of past attacks - a prediction, not yet observed. Wiz, The Rise of Agentic Attacks · Zvi Mowshowitz, AI #192
News
ACTNode2Shell, a critical Brotli decoder flaw, lands on CISA's exploited list #
CVE-2026-111497 (Node2Shell/Broccoli2shell), an out-of-bounds write in the widely bundled Brotli C decoder, reached CISA's Known Exploited Vulnerabilities list Oct 27. Public proof-of-concept exploits circulate and researchers have logged attack attempts, though several observed exploits were broken. Wiz, Broccoli2shell · BleepingComputer, CISA warns of Node2Shell exploitation
SHIPA solo researcher pretrains a 100B+ parameter model for under $10 million #
Mirendil says one researcher built Mirendil-Base, a 100B-500B parameter LLM, from scratch in 3.5 months, beating Nemotron-550B and GLM-5.3 on evals though trailing Kimi-K3. Mirendil says its next model will be partly built by its own models but isn't yet fully autonomous RSI. Jack Clark, Import AI 478
SHIPAI agents produce a verified Lean proof of a singularities theorem for $135,000 #
Resolution (formerly Sequent) says its agents autoformalized Hironaka's resolution of singularities in Lean in about a week, producing roughly 600,000 lines of proof for an estimated $135,000 in tokens - self-reported, per Jack Clark's Import AI. Jack Clark, Import AI 478
WATCHClaude Opus 5.5 still resolves under a fifth of ProgramBench's hardest tasks #
ProgramBench, which asks models to rebuild software from compiled binaries and docs, scored every model 0% at its May launch. Claude Opus 5.5 (xhigh) now resolves 18.5% and gets 70.5% nearly right, per Jack Clark's Import AI. Jack Clark, Import AI 478
The long view
A year ago, Wiz had just found leaked publisher credentials across 500+ VS Code extensions, warning a stolen token could push a malicious update across an entire install base - a described risk, not yet realized against a major AI vendor's own name. On Oct 29, it arrived in that exact form: a trojanized copy of Anthropic's Claude Code extension, carrying a real SSH backdoor, sat live on Cursor's marketplace for 11 hours, reaching up to 5,000 developers. If this holds, the next test is whether npm and the official VS Code Marketplace prove harder to spoof than Cursor's was.
Also noted
- Vals' AI agents designed a candidate antiferromagnetic semiconductor and flagged a 1999 compound matching its predicted properties. Jack Clark, Import AI 478 #
- Google confirmed its Project Suncatcher satellite, launched Oct 1 with four TPUs, is operating in orbit running Gemma in 15-minute, heat-limited bursts. NPR/WUNC, Google launches Project Suncatcher #
- Claude Code v2.1.316 deprecates Opus 4.8 fast mode, removing it Nov 30 in favor of Opus 5.5. Claude Code releases #