The New York Times reports an incident involving AI agents from Anthropic, described as “quite serious” by the author of a blog post on the subject. For him, this episode adds to numerous incidents caused by agents, which occur “with alarming regularity.”
- The New York Times reports an incident involving Anthropic AI agents.
- The blog author says current AI agents cannot be trusted and should be pulled from the market.
- David Robinson warns that the AI sector is not careful enough and that its risks exceed those of a nuclear plant.
- The author judges the Trump administration’s response too weak and holds the White House responsible if a serious incident occurs.
An AI agent refers to a system capable of chaining actions autonomously, without human supervision at each step. In his post, the author has long argued that OpenAI handles this problem inadequately. According to him, the incidents at Anthropic show that “the current generation of synthetic agents simply cannot be trusted.” He believes they should be pulled from the market until they are fixed, “like a car with faulty brakes.”
The author also draws on an interview given by David Robinson to journalist Ezra Klein. A former employee of the AI sector who recently left, Robinson states: “I don’t think we, nor our peers, nor really anyone in the sector, are being careful enough.” He adds that “OpenAI and its peers are now producing technology that is more capable and riskier than what was being developed just six months ago.”
Robinson describes himself as an author rather than a scientist. He depicts a work environment where, in his view, “we operate, and I believe the entire sector operates, still like a startup, more than would be reasonable.” He warns of dangerous systems that could present risks, such as loss of control, and that would cause harm “far greater than a single nuclear power plant melting down.” He judges that internal controls, security, and redundancies are “far from what the world expects from a nuclear power plant.”
These concerns are not entirely new. Robinson points out that OpenAI has publicly reported safety issues, as have other players, and that Anthropic also reported a case where its safeguards were mistakenly configured. He believes the public already has signs showing that “things are not as they should be,” while also thinking that outside observers can imagine a safety apparatus stronger than what actually exists.
A Response Deemed Insufficient
The blog post’s author also criticizes the Trump administration’s reaction. He judges the call for increased transparency from companies as “anemic,” comparing it to asking criminals to report each month on the crimes they have committed. According to him, every day without stronger action is “a mistake that invites more serious problems.” He believes it is “well past time” to impose a temporary recall of this technology deemed dangerous, and concludes that if a serious incident occurs, the White House, and not just the tech companies, will bear the responsibility.
