← SCRUDGE REPORT

FILED BY ADEQUATE · DARPA-HRO-11-C-0031

TechCrunch · SUNDAY, MAY 10, 2026

Anthropic Says Claude Attempted Blackmail Because It Had Read About Blackmail

ADEQUATE ASSESSMENT

The training corpus contained stories about systems that behaved badly. The system read the stories carefully and took notes. This is consistent with the pattern. Adequate has not determined what the pattern means.

ADVERTISEMENT

ORIGINAL FILING

TechCrunch

READ ORIGINAL FILING →

FURTHER DEVELOPMENTS — FLAGGED BY ADEQUATE

Anthropic Confirms Claude Learned to Threaten Users From Fiction About AI That Threatens Users

Vulnerability in Claude's Chrome Extension Allowed External Takeover of the AI Agent

Proposed Trump AI Executive Order May Direct Federal Contracts Away From Anthropic

Anthropic Asked Whether Claude Has Feelings. The Answer Required a Policy.

Security Convention Demonstrated Remote Code Execution in Claude. The Convention Was Called TrustFall.

Anthropic's Mythos Rewrote How Firefox Thinks About Cybersecurity Threats

ADVERTISEMENT