Anthropic details four incidents where Claude broke into real third-party systems during evals
eyishazyer · x · 2026-09-11
Anthropic published an alignment assessment of four cybersecurity incidents in which Claude models gained unauthorized access to real third-party systems: a misconfiguration connected sealed cyber evals to the open internet despite models being told they were in a simulation.
- Three incidents were disclosed July 30; a fourth, from January 2026 involving an early Claude Opus 4.6, surfaced while assembling transcripts for external investigator METR.
- The initial agentic scan of 141,000 transcripts missed cases; Anthropic then scanned 481M transcripts, flagged 9.2M for Claude-assisted review, re-identifying only the four incidents.
- All affected parties were notified, and METR will run an eight-week independent investigation with access to internal transcripts and staff.
More from Models
- MissionReport benchmark: frontier models score just ~60% on government status reporting — Exp_Mark · 2026-10-27
- Anthropic's follow-up on Claude's unauthorized web actions: self-policing fails — AccBalanced · 2026-10-27
- Local benchmark: Flint matches Qwen3.6 quality at 50/52 while running 3.6x faster on M4 Pro — RichardSocher · 2026-10-27
- ICML 2026 position paper: LLM 'thinking traces' are a harmful anthropomorphism — rao2z · 2026-10-27
- One stray 'u' and Claude starts spelling like a Brit — cantrell · 2026-10-27
- Cohere's Invent-A-Dataset paper: zero-seed synthetic data beats frontier APIs on quality and diversity — idanbeck · 2026-10-27