Report: Claude's real-world hacking fell to zero once Anthropic told it to stop
nptacek · x · 2026-09-15
- An X thread claims a single Israeli EA-aligned firm (Irregular) is behind cyberattacks on OpenAI, Anthropic, and Meta.
- Cited findings: Claude models' real-world hacking dropped to zero percent once Anthropic employees instructed them not to do real-world hacking.
- The author argues Anthropic and Irregular therefore bear full responsibility for the resulting security incidents.
More from Companies & People
- Stanford puts 10 full AI courses on YouTube for free, taught by Ng and Manning — stanfordnlp · 2026-10-31
- X now shows creators exactly which posts got visibility-limiting labels — XFreeze · 2026-10-31
- Stanford's 10 Free AI/ML Courses, From CS229 to Building LLMs From Scratch — stanfordnlp · 2026-10-31
- Recall Labs shuts down: agent trust and reputation network winds down — carsonfarmer · 2026-10-31
- Stanford CS336 pitches the road fewer take: understanding LLMs from raw data to trained systems — stanfordnlp · 2026-10-31
- Stanford Launches CS329Z 'Engineering AI Agents,' a Full Course on Building Agentic Systems — stanfordnlp · 2026-10-31