Two companies that compete directly for customers, talent, and market position were, earlier this year, in talks to poke holes in each other’s most sensitive technology. According to the New York Post, OpenAI and Anthropic held negotiations on a legally binding agreement that would have required each company to run the other’s new models through rigorous safety tests, searching for flaws and what the report calls “hidden dangers.” Lawyers were apparently already drafting terms. Whether the deal was ever finalized is still unclear.
The source is The Information, citing someone with direct knowledge of the talks. That’s a credible enough origin to take seriously, and the timing matters: negotiations reportedly began before OpenAI was linked to an accidental hack of Hugging Face, which means this wasn’t a reaction to bad press. It was something both sides saw value in before things got messy publicly.
So why does this matter? Because it signals that even the labs themselves aren’t fully confident in their own internal safety evaluations. If OpenAI and Anthropic, two organizations with enormous safety research teams and significant public accountability pressure, were willing to let a competitor stress-test their models, that’s an implicit admission that outside scrutiny has real value. For developers and founders building on top of these platforms, that’s worth paying attention to.
The broader context here is a safety debate that has accelerated fast. Anthropic CEO Dario Amodei recently published an essay calling for an industry slowdown, warning that AI systems could be close to causing catastrophic damage without stronger guardrails. A former employee of both Anthropic and OpenAI said publicly that leading labs are “gambling with our lives.” Elon Musk, whose xAI builds the Grok models, went further at the All-In Summit and argued that top US and Chinese labs should be testing each other’s models directly.
But the safety conversation is also getting complicated by questions of independence. Amodei has endorsed METR, an evaluator group with financial ties to Anthropic’s own investors and deep connections to the Effective Altruism movement, as a model for embedded third-party oversight. Critics say that’s not independence, it’s a conflict of interest dressed up as accountability. Redwood Research has similar ties. Sam Altman has said he supports the idea of third-party watchdogs in principle, without backing any specific group.
On the other side, President Trump has called AI safety concerns a “hoax” and announced plans to create an “AI Force” with a dedicated czar to oversee the industry, suggesting the regulatory picture at the federal level remains unpredictable.
For anyone building AI products or evaluating which foundation model to build on, the OpenAI-Anthropic talks are a signal worth tracking. Peer review between competitors, if it ever becomes standard practice, would change how model risk is communicated externally and how enterprise buyers assess trust. It hasn’t happened yet. But the fact that lawyers were already involved suggests it came closer than most people knew.
