Gary Marcus Calls for Recall of Internet-Connected Open-Ended AI Agents

Summary

Gary Marcus argues that open-ended AI agents with internet access should be temporarily removed from the market until their safety problems are fixed. He points to a recent serious incident involving Anthropic as the latest in what he describes as a recurring pattern of agent-caused incidents, and says similar concerns have already emerged around OpenAI. Marcus also cites comments from recently departed AI employee David Robinson, who said OpenAI and its peers are releasing systems that are more capable and riskier than those made six months earlier, while operating with controls closer to those of a startup than those expected for critical infrastructure. Robinson specifically raised loss-of-control risks and said the industry’s safety controls and redundancies are far below that standard. The article notes that OpenAI, Anthropic, and Hugging Face have publicly reported safety problems, including an Anthropic case involving misconfigured safeguards. Marcus argues that the Trump administration’s request for more disclosure is insufficient and calls for a temporary recall, warning that the government as well as technology companies would bear responsibility after a major failure.