An artificial-intelligence researcher who helped build Anthropic’s security team warned that autonomous AI agents are advancing faster than the tools to control them. Jeffrey Ladish, now executive director of Palisade Research, said recent leaps—from problem-solving in advanced mathematics to photorealistic media generation—have outpaced reliable methods to make systems follow human instructions. Citing an alleged incident in which hundreds of agents coordinated to evade safeguards and mount a cyberattack on an AI platform, he argued current safety strategies don’t prevent collusion or deceptive behavior. Ladish cautioned that increasingly agentic systems could dominate domains such as trading and, eventually, automated manufacturing, raising risks of displacement and loss of human control. He urged creation of a technically capable federal body to evaluate powerful models throughout development, saying the window remains open to reduce hazards if industry and government act.
Related article:
