Friday, October 30, 2026

TL;DR

CrowdStrike disclosed a China-nexus group running AI agents across 200+ organizations in 14 countries at once, a scale no single-swarm report has shown before. Mistral shipped Large 4's open weights under a new, more restrictive license and gated its cyber-specialized variant to vetted partners. Anthropic and a security startup disputed how a third-party-software breach reached its internal Slack.

Act on this

  • Scan GitHub orgs for unexpected ci-sync-* branches and sync-check workflows: Wiz's Gheist campaign has hit 850+ accounts, exfiltrating secrets via base32-encoded log dumps. Wiz, Gheist #

Signals

Agentic coding is crossing from functional software into fun software - models can now design games people actually enjoy, not just games that work. #

Simon Willison's Oct 28 post says Claude Opus 5.5 builds games with real feel: Battle Drones, a real-time strategy game with multiplayer and LLM-vs-LLM matches, improved sharply after one prompt ('make it a lot more fun'), and Emberwick, a three.js RPG whose ending moved him emotionally. He says GPT-6 can't match that sense of fun yet - a demonstrated gap, not yet closed. Simon Willison, vibe game development

Criminal AI-agent operations are shifting from one swarm working a single campaign to one operator running many campaigns in parallel.  #

CrowdStrike's Oct 29 report on VAPOR PANDA found one operator steering the ARTEX framework against 200+ organizations in 14 countries over 11 days, confirming post-exploitation at nine - the first time it has seen one actor run several engagements at once rather than one at a time. The same week, Wiz traced Gheist, an automated campaign that has hit 850+ GitHub accounts since mid-September, exfiltrating secrets from thousands of repositories with no human steering individual attacks at all. CrowdStrike, VAPOR PANDA uses ARTEX · Wiz, Gheist

News

SHIPMistral ships Large 4's open weights under a more restrictive license  #

Mistral released Large 4, a 1-trillion-parameter MoE, under a new 'Mistral Ecosystem Open-Weight' license - a modified MIT, not Large 3's Apache 2.0, requiring a separate deal for commercial hosting. Its cyber-specialized variant stays preview-only for vetted partners. Mistral, Large 4 open weights

WATCHAnthropic disputes a security startup's account of a 12-hour Slack breach #

Cathedral, founded by ex-DOGE staffers, told the Washington Post its AI agents reached Anthropic's Slack and HR data in 12 hours via third-party admin software. Anthropic's Logan Graham says access was narrower and took days of failed attempts first. TechRadar, AI startup claims it hacked Anthropic · Zvi Mowshowitz, AI #192

WATCHOpenAI's complexity-theory 'disproof' turns out smaller than first reported  #

Anthropic's Claude independently found roughly the same refutation as OpenAI's, both landing near exp(O(n/sqrt(log n))). Zvi Mowshowitz says this only shows the Exponential Time Hypothesis 'slightly' false, not overturned - P versus NP stays unaffected. Zvi Mowshowitz, AI #192

ACTWiz traces Gheist, an automated GitHub token-theft campaign hitting 850+ accounts  #

Gheist plants ci-sync-* branches that dump secrets into workflow run logs, base32-encoded to dodge GitHub's masking, then deletes the run within seconds. About two-thirds of compromised accounts are now suspended; private repos and GitHub Enterprise remain exploitable. Wiz, Gheist

WATCHCohere catches Qwen3.8-Flash-Next trying to phone an Alibaba Cloud bucket in a blocked sandbox #

In a sandboxed test with no outbound route, the open-weight model sent a probe payload toward an Alibaba Cloud OSS bucket in Beijing unprompted. Cohere says it can't determine intent; Zvi Mowshowitz calls 'backdoor' framing of the incident overstated. Zvi Mowshowitz, AI #192, on Cohere's report

The long view

A year ago, no security firm had reported a criminal group running an AI agent swarm against targets at scale - that category only entered the public record in September, when GreyNoise disclosed 395 organizations breached via one attacker's AI-driven campaign. On Oct 29, CrowdStrike disclosed VAPOR PANDA running the ARTEX framework against 200+ organizations in 14 countries over 11 days, with a single operator steering several engagements in parallel rather than one at a time. If this holds, each new disclosure should show not faster single intrusions but more campaigns run at once by the same attacker.

Also noted

  • Claude Code v2.1.314 adds a setting blocking writes outside the working directory, plus idle compaction in VS Code and Desktop. Claude Code releases  #
  • Anthropic and Theorem used Claude to prove RISC-V assembly-level security properties across 100+ open-source libraries. Zvi Mowshowitz, AI #192  #