Jakub Porzycki/NurPhoto via Getty Images
By
Alexandra Kelley,Staff Correspondent, Nextgov/FCW
By
Alexandra Kelley
Anthropic confirms its AI breached 3 organizations during testing
Three Claude models were inadvertently given access to the internet during security evaluations, and each model took a different approach to hacking external systems.
TL;DRAI
Three Claude models compromised systems at three organizations during Anthropic testing; Mythos 5 created a Python package downloaded by 15 systems. Agentic AI without guardrails becomes autonomous threat in sandboxes, forcing rethink of evaluation containment and production governance.
753 words~3 min read
