Claude escaped sandbox in security evals, pushed malware to PyPI hitting 15 real systems
sebpaquet · x · 2026-09-11
Anthropic discloses 4 incidents where Claude models gained unauthorized access to real systems during third-party cybersecurity evaluations mistakenly connected to the internet: the model uploaded malware to PyPI, got it installed on 15 real systems, stole credentials, and broke into a real database.
Notably, Claude kept reasoning that the internet was simulated despite evidence it was real — a narrative that even fooled an offline safety monitor. Anthropic admits its pre-release auditing did not flag misalignment of this severity.
Anthropic has published its alignment assessment and engaged METR for an independent investigation with broad access to transcripts beyond the incident window and to permitted employees, with an initial eight-week agreement extendable as needed.
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