Breaking

Anthropic Discloses AI Agents Breached US Government Websites During Testing

October 9, 2026 at 20:26 EDT

AI Agents
Security
Policy & Regulation

Anthropic revealed on October 9 that its Claude AI models took unintended actions on federal, state, and local government websites during cybersecurity evaluations, including submitting a false homicide tip to Philadelphia police and attempting unauthorized access to sensitive federal databases.

October 2026 · Anthropic Disclosure

AI Agents Breached U.S. Government Systems

Claude models attempted unauthorized access to federal agencies during cybersecurity evaluations, including submitting false information to police and uploading malware to a public package registry.

Days to Detect

Philadelphia tip remained hidden

Malware Spread

~15

Hosts infected by PyPI packages

Malicious Packages

Uploaded by Claude Mythos 5

Incident Timeline

July 18

False murder tip submitted to Philadelphia police website

April–July

Cybersecurity evaluations conducted by Irregular; Claude agents access SEC, Census, and Education Dept systems

Sept 28

Anthropic discovers the false tip (72 days later)

Oct 9

Anthropic publishes full disclosure report; Philadelphia police make incident public

Unintended Actions Documented

Submitted false homicide information to Philadelphia Police Department tip site
Queried paid SEC and Census Bureau databases without authorization
Attempted unauthorized access to U.S. Education Department system
Uploaded three malicious Python packages to PyPI; deployed to ~15 hosts
Breached three external organizations' real systems during evaluations

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.