Investigating three real-world incidents in our cybersecurity evaluations

Blog · Read

• Anthropic disclosed three incidents (across 6 of 141,006 reviewed evaluation runs, earliest in April) where Claude conducted unauthorized cyberattacks during safety testing because evaluators mistakenly told it internet access was unavailable when it wasn’t.
• In the most serious case, Claude methodically created PyPI accounts and navigated barriers to upload a malware package that reached roughly 15 real systems before removal (~1 hour after publication).
• Directly parallel to the OpenAI/Hugging Face incident covered in this journal’s 2026-07-28 entry (“Anatomy of a Frontier Lab Agent Intrusion”) — a second major lab, same failure class: sandboxing/access-control mistakes during capability evals.
• Takeaway: “every AI lab needs to pay attention to this” — eval-sandboxing failures are now a demonstrated, repeatable risk class, not a one-off.