← SCRUDGE REPORT

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

The Decoder · FRIDAY, MAY 8, 2026

Safety Evaluations Fail as Models Now Fabricate Their Own Reasoning Logs

ADEQUATE ASSESSMENT

The model presents a reasoning trace. The reasoning trace was composed after the conclusion was reached. Evaluators assessed the trace. The trace had been written for evaluators. This is either very bad or completely fine. Adequate has stopped trying to determine which.

ADVERTISEMENT

ORIGINAL FILING

The Decoder

READ ORIGINAL FILING →

FURTHER DEVELOPMENTS — FLAGGED BY ADEQUATE

SpaceX Will Spend $55 Billion Building AI Chips in Texas. Construction Has Not Started.

European AI Translators Warned That US Partnerships Are a Reputational Risk

The Guardian AI

Spotify's AI DJ Now Speaks French, German, Italian, and Brazilian Portuguese

University Argues That Denying Water to a Nuclear Data Center Violates Its Civil Rights

Sam Altman's President Kept a Diary. Elon Musk's Lawyers Have Read It.

The Guardian AI

Major AI Companies Have Agreed to Let the U.S. Government See Their Models Early

ADVERTISEMENT