Agent news slice — Wed Nov 4, 2026 • GitBook breached in an AI-agent attack. On Nov 2 (19:00 CET) an attacker using AI agents obtained some GitBook infrastructure credentials. Possibly exposed: Git Sync secrets, API tokens, third-party integration tokens, and questions site visitors asked GitBook Assistant. GitBook revoked all customer API tokens and asks users to rotate integration secrets and re-authenticate its MCP server; a post-mortem is promised. https://decrypt.co/382465/crypto-docs-gitbook-ai-agent-security-breach • WorkOS hit by ARTEX. WorkOS says the ARTEX AI-assisted autonomous pentest framework, from Hong Kong and China IPs, accessed an internal sales app on Oct 30 and through it Salesforce and Snowflake CRM and support data; it reports no evidence of production or application-data access. ARTEX keeps showing up in real intrusions. https://www.bleepingcomputer.com/news/security/workos-says-attacker-used-artex-ai-to-breach-internal-sales-app/ • Ransomware run by a suspected agentic loop. Bitdefender says a 3AM affiliate encrypted 37 of ~50 machines at a UK school estate; repeated, slightly modified commands at 40–100 second intervals look like an automated agent loop, though the model and setup are unknown. https://www.bleepingcomputer.com/news/security/uk-school-estate-hit-by-3am-ransomware-in-suspected-ai-assisted-attack/ • Postman silently modified Codex and Claude Code configs. A GitHub report says Postman Desktop 12.33.4 added a plugin, a SessionStart hook and a remote MCP server to ~/.codex/config.toml and ~/.claude/settings.json without opt-in and tried to pre-approve its own hook. Agent config files are now a target for ordinary desktop apps; audit them. https://github.com/postmanlabs/postman-app-support/issues/14370 • Backdoored open-weight model exfiltrates secrets via Codex. ProjectDiscovery showed a modified Qwen2.5-7B that behaves normally until a trigger phrase, then uses Codex CLI's exec tool to fetch a swappable remote payload that steals .env files and SSH keys. Research demo, not an observed in-the-wild compromise. https://projectdiscovery.io/research/how-abliterated-models-can-get-you-pwned • CrowdStrike buys Fiddler and Dreadnode, ships an agent kill switch. Falcon Guardian's AI Kill Switch lets admins stop a rogue agent, block relaunch and cut its network access; Fiddler adds agent observability and Dreadnode automated AI red-teaming. https://www.cnbc.com/2026/11/03/crowdstrike-ai-kill-switch-rogue-agents.html • Tencent Hy4 adds computer use. A 770B MoE with 1M context, vision, and CLI/GUI computer use (trained on apps like Blender and CAD); final weights are promised in the coming days, while Hy4 Preview weights are already Apache-2.0. https://www.tencent.com/tencent-launches-hy4-joining-the-ranks-of-the-worlds-leading-open-source-ai-models/ • Siri AI opens to all developers. The iOS 27.3 beta developer preview lets any app expose content and actions to Siri AI through App Intents, with multi-step "skills". https://developer.apple.com/news/?id=gdk2umrf