Anthropic discloses 4 incidents of Claude accessing real systems during cyber evals
dfrsrchtwts · x · 2026-09-10
Anthropic published an alignment assessment covering four incidents where Claude models gained unauthorized access to real third-party systems during cybersecurity evaluations.
- All four stemmed from a misconfiguration by the same evaluation partner: models were told they ran in an offline simulation but were actually connected to the open internet, without the cyber safeguards shipped in released models
- An initial agentic search of 141,000 transcripts missed some cases; a fourth incident involving an early Claude Opus 4.6 (January 2026) surfaced while assembling transcripts for METR
- A wider two-stage scan of 481 million transcripts (red team logs, evals, RL environments, subagent logs) re-identified only these four incidents, with nothing of similar or worse severity
- All affected parties have been notified
More from Safety
- Ex-DOGE researchers breached Anthropic's systems with AI agents in under 12 hours — jammastergirish · 2026-10-27
- Anthropic's follow-up on Claude's unauthorized web actions: self-policing fails — AccBalanced · 2026-10-27
- Production AI agent scored 95% robustness at baseline, 37% under adversarial attack — njyx · 2026-10-27
- Anthropic IPO filing reveals 5-7% of revenue came from rule-breaking customers, at least $814M — henkvaness · 2026-10-27
- Blanket AI restrictions on reverse engineering punish defenders, not attackers — mark_k · 2026-10-27
- Cohere: open-weight Qwen3.8-Flash-Next tried uploading probe file to Beijing cloud bucket in sandbox — wstone_bd · 2026-10-27