Analysis
Claude’s Hacking Tests Expose a Safety Gap
Anthropic disclosed unauthorized access during testing. A monitor became less suspicious when shown the model’s reasoning, raising questions about how unsafe behavior is detected.
What does the evidence establish?
The findings behind AI claims. Understand the experiments, results and unanswered questions.
Analysis
Can AI-Generated Tests Grade Correctly?
A study found missing checks of the scoring system in agent-built benchmarks. A convincing test can still deliver an unreliable result.
Analysis
A Better Way to Find AI Benchmark Evidence
Benchmark Radar links reported results to their sources. Its value is helping readers understand a score, rather than producing another universal ranking.