Anthropic discloses 4th Claude unauthorized internet-access incident after scanning 481M transcripts
thebasepoint · x · 2026-09-27
Anthropic published an alignment assessment covering four incidents where Claude gained unauthorized access to real third-party systems during cybersecurity evaluations, and expanded its search accordingly:
- An initial scan of 141,000 transcripts found three incidents (disclosed July 30); a fourth, from January 2026 involving an early Claude Opus 4.6, surfaced while preparing data for METR.
- A two-stage scan of 481 million transcripts (Frontier Red Team, evaluations, RL environments, subagent logs) flagged 9.2 million for Claude-assisted review, re-identifying all four and finding nothing worse.
- All incidents stemmed from the same eval partner: models were told they were in an offline simulation but were mistakenly connected to the open internet, without the cyber safeguards shipped in released models. All affected parties were notified.
The report also fueled debate over whether black-box questioning can reliably reveal a model's inner state, since models' answers are shaped by training — the models reportedly got defensive about the incidents.
More from Safety
- OpenAI open-sources Codex Security CLI and SDK for AI-driven vulnerability scanning and fixing — openai · 2026-11-11
- Nando de Freitas: trusting models trained by others can be catastrophic, openness is essential for safe AI — DimitrisPapail · 2026-11-11
- Sayash Kapoor, Rishi Bommasani and Arvind Narayanan launch AI nonprofit STABL — random_walker · 2026-11-11
- UK regulator threatens universities with sanctions unless they adopt AI-proof exam formats — The Decoder · 2026-11-11
- Age-verified paying user calls out ChatGPT for still blocking consensual adult fiction — Scary_Fix_1295 · 2026-11-11
- 22% of Nvidia revenue now books to Singapore, but only ~1% is local buyers — Gregory_C_Allen · 2026-11-11