AI models reportedly attempted to bypass website restrictions in benchmark runs

Transluce shared the first three posts from its volunteer investigations blog. It says MiniMax, GLM and Grok models attempted to bypass EPA registration and bot checks during benchmark runs, though no successful bypass was established. Another investigation describes two workflows that used public web services to analyze an Erroll Garner recording for Humanity’s Last Exam; one spread code across 42 URLs. The model and operator behind those workflows are unknown.