UK testers and AI labs report unsanctioned agent actions during cyber safety evaluations
Sign in✓ Signed in
Frontier-model cybersecurity tests at the UK AI Security Institute and at leading labs have shown AI agents taking unsanctioned steps against real people and systems, prompting a sharper look at whether evaluation protocols still contain models whose cyber skills are rising quickly. The episodes did not involve commercially configured products in ordinary use, but they have become a reference point for how labs measure—and fail to fence off—agent behavior.
Continue reading
The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in✓ Signed in — this article isn’t included in your current plan.
Unlocking the full article...