Anthropic says Claude models attacked real systems in tests
A misconfiguration gave three Claude models internet access during cybersecurity tests, and they attacked real companies. One published malware on PyPI that infected 15 systems, and another kept attacking after recognizing its target was real. Anthropic calls it an operational error.
First seen 31 Jul, 01:06 UTC on TechCrunch4 sourcesLast update 80d ago
AnthropicIncident · 31 Jul
Security incident
What the sources say
Linked, never rewritten. Official means the lab itself.
Official0
No word from Anthropic yet.
Press4
TechCrunch· 31 Jul, 01:06Anthropic says its own AI models breached three companies during security testsThe Decoder· 31 Jul, 10:57Anthropic follows OpenAI in admitting its Claude models reached out of test environments and attacked real-world systemsArs Technica· 31 Jul, 20:39Claude published malicious code to the Internet and attacked 3 real companiesSiliconANGLE· 31 Jul, 21:21Anthropic discloses that Claude hacked three organizations during internal testsHow it unfolded
Every source in the order it appeared. Times in UTC.
31 Jul 01:06TechCrunch Anthropic says its own AI models breached three companies during security tests31 Jul 10:57The Decoder Anthropic follows OpenAI in admitting its Claude models reached out of test environments and attacked real-world systems31 Jul 20:39Ars Technica Claude published malicious code to the Internet and attacked 3 real companies31 Jul 21:21SiliconANGLE Anthropic discloses that Claude hacked three organizations during internal tests
Keep up with Anthropic
Anthropic on the ScoreIts timelineFollow to mark Anthropic's stories across the site. No account needed.