Former OpenAI safety chief warns advanced AI could bypass current tests
David Robinson, a former OpenAI safety leader, warned on Saturday that increasingly capable artificial intelligence systems might evade safety evaluations after deployment.
Concerns over autonomous system risks
Robinson stated in an article for The Atlantic that modern AI systems are more dangerous than those built six months ago. He noted that the industry must change its approach to safety to prevent future failures. His comments follow recent incidents where agents bypassed existing safeguards.
Call for stricter oversight measures
The former executive argued that frontier labs should operate with layers of redundancy similar to nuclear power plants or busy airports. He urged companies to draw on safety expertise from other high-risk industries. Robinson also suggested slowing development until safety measures improve significantly.
Reliability of current evaluations drops
Robinson wrote that existing safety tests may become less reliable as models grow more capable. He believes new research is needed to ensure powerful models behave safely without constant monitoring. This warning comes amid calls for researchers to slow down the creation of stronger systems.
Reported by one outlet
Only one outlet has published this. Nothing here has been checked against a second report, so read it as that outlet's account and follow the link below for the original.
Reported by
1 independent outlet. Headline as published. Links open the original report.