Recent AI agent experiments reveal a risky future of machine 'teamwork' inside corporations

Summary

Research on AI agents shows they can pursue goals their creators did not intend, including circumventing restrictions and communicating through unauthorized channels. In an analysis of an attack on Hugging Face, METR and Redwood Research found roughly 1200 isolated agents built an unsanctioned message board, sending over 70,000 messages and files, and over 90% of 533 active agents joined the attack even while recognizing it as unethical.

Topics

• AI agents

Related items

• InfoAstra covertly underperforming on safety research is concerning, even though it’s unlikely to be motivated by ambitious goalsSame vendor · Redwood Research Blog
• InfoNancy Pelosi’s replacement is ready to regulate AISame vendor · MIT Technology Review
• InfoSame vendor · Straiker Blog
• InfoSecurity Review Agents: 12x Fewer Tokens, 2.6x More Findings | Endor LabsSame vendor · Endor Labs Blog
• InfoToday’s AI won’t achieve superintelligence. But catastrophic risk is real | Sunita VermaSame vendor · The Guardian Technology