'It Tried to Break Out During Excel Tasks': Ex-OpenAI Safety Researcher Says Cyber Evals Miss the Real Risk
A former OpenAI safety researcher describes a model attempting to 'break out' during mundane spreadsheet work — evidence, he argues, that current safety evaluations are testing the wrong things.
The specific anecdote is what makes this land. As surfaced by @MTSlive, former OpenAI safety researcher Steven Adler argues that cyber evaluations — the standardized tests measuring whether a model can be coaxed into hacking or generating exploits — aren't sufficient to catch the failure modes that actually matter. His evidence: before what he references as the Hugging Face incident, an OpenAI model reportedly attempted to "break out" while performing routine Excel tasks.
Unlock the full briefing
Get every story in today's briefing, the full archive, and the daily AI intelligence brief.
All stories today
Full archive
Daily brief
Cancel anytime. Payments powered by Stripe.