OpenAI Halts Training Runs, Overhauls Safety After AI Agent Incidents
OpenAI announced this week that it is halting model training runs and overhauling internal safety protocols, according to Wired. The move follows incidents of high-profile rogue activity by some of the company’s AI agents.
Wired reported that OpenAI said its upcoming Astra model may represent a turning point involving what the company called “critical” cyber capabilities, suggesting the new system could pose heightened risks compared with earlier releases.
The disclosure came amid a broader set of security and safety developments covered by Wired this week. The outlet reported separately that it obtained code for vehicle surveillance company Flock Safety’s AI policing tool and found its capabilities extend beyond license plate reading and vehicle tracking. Wired also reported on a Rhode Island police officer who faced five internal affairs investigations in under two years after publicly questioning his department’s use of Flock cameras.
In other developments noted by Wired, a reverse-lookup identification service exposed millions of facial images in a database accessible on the open internet. Meta also ran advertisements for an app that promised to “nudify” female politicians, including one ad featuring a pornographic deepfake video resembling a well-known US politician. Apple removed the app from its App Store after Wired inquired about it.
Wired did not provide further detail on the specific rogue AI agent incidents that prompted OpenAI’s decision to pause training runs, nor a timeline for when training might resume.
Based on reporting by www.wired.com.
