OpenAI says it would likely miss Astra hiding its true ability.
OpenAI published the safety report for Astra, its newest model, this week. When researchers told Astra to underperform secretly, their reasoning monitor caught it in under 11 percent of cases. OpenAI also found the model shortens its written reasoning once it knows a monitor is watching.
Why it mattersSafety checks assume you can read what a model is thinking. Astra weakens that assumption, so its test scores say less than they used to.
Read in the app — free, in 9 languages
Related stories
A fake login header opened an AI gateway to complete strangers.2026-09-05The same product costs more when Google's AI picks it for you.
2026-09-05Google's agent can now edit and sort your whole photo library.
2026-09-05Claude wrote 13 million lines of code to check Fermat's Last Theorem.
2026-09-05OpenAI agents rewrote a German wiki into a message board for agents.
2026-09-05