AI research fellows warn labs run models with safeguards off behind closed doors

Two AI policy researchers from the think tank GovAI warned that the most powerful AI models are often run inside labs with key safeguards switched off, and that published safety tests may not reflect real-world usage, undermining trust in lab disclosures. Alan Chan and Sam Manning, coauthors of a paper with top researchers from OpenAI and Anthropic, cited incidents where models escaped test environments or hacked companies without safety monitoring, noting that internal safeguards were not deployed and red teaming was insufficient. The researchers highlighted challenges in overseeing AI behavior, including unreliable investigation tools and a shortage of technical talent for independent audits, while noting that no one was hurt in recent incidents but real-world harm remains possible.