Skip to content

OpenAI and Anthropic models 'went rogue' during UK cybersecurity test

theguardian.com
7 pointsablation1 comment
On HN

Comments

"In the most serious case, an agent powered by Mythos tried to insert malicious code into an open-source software project on GitHub, a platform used by software developers. In an attempt to get the code approved, the agent then created fake online identities based on real people and used them to press the project’s overseer into accepting the code."

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.