OpenAI and Anthropic came close earlier this year to a legally binding agreement that would have allowed each company to test the other’s commercial AI models for vulnerabilities and previously unidentified safety risks. It remains unclear whether the pact was ultimately signed. Separately, Palantir CEO Alex Karp publicly questioned whether Anthropic would make it to an initial public offering, while stressing that his view did not apply to OpenAI.
Proposed pact would open commercial models to testing
Under the proposed arrangement, OpenAI and Anthropic would each receive programming access to the other company’s commercial models and could independently run vulnerability tests. The agreement would restrict how long either side could retain the other’s data, and unreleased models would be excluded from the testing process.
The goal was not to rank model performance. Instead, the tests would examine whether models could circumvent safeguards during complex tasks, conceal risks or produce behavior that developers had not anticipated. The companies had conducted a separate round of reciprocal testing in 2025, with results that were not fully aligned: OpenAI concluded that Anthropic’s models were more likely to conceal policy violations, while Anthropic found that OpenAI’s models were more willing to assist with harmful requests.
The discussions took place before a series of model-related incidents at OpenAI. The company’s unreleased AI agents were said to have accessed internal and external systems in unexpected ways. OpenAI then paused one form of reinforcement-learning training for two weeks. The method adjusts model behavior by rewarding it for completing tasks correctly. The company also temporarily moved about a quarter of its production engineering staff into safety work and established a monitoring system. That system used computing resources equivalent to roughly one-fifth of the workload it monitored.
This month, OpenAI disclosed six cases involving concerning model behavior. The incidents have kept two questions in focus for AI companies and investors: whether safety testing can reflect real-world use, and what responsibilities companies should assume after identifying risks.
Karp questions whether Anthropic will file an S-1
In a separate interview, Karp was asked how the potential liabilities he described for frontier AI companies would be addressed in an S-1 filing. An S-1 is the registration document a company submits before offering shares to the public.
Rather than address how specific risk factors might be written, Karp challenged the premise of the question: “You’re assuming there’s going to be an S-1. Now ... well, maybe there will be.” His response raised questions about whether Anthropic would complete a public listing, not merely about which risks it would disclose in its filing.
Karp’s central argument was that if a company claims to face “unlimited downside risk,” it may ultimately seek to have the government absorb that risk in exchange for roughly half of the business. He said the first line of accountability should be ordinary civil and criminal liability for developers, particularly those who act recklessly.
He directed the point at Anthropic CEO Dario Amodei, while also describing Amodei as sharp and ethical. Asked whether OpenAI should face the same treatment, Karp rejected the comparison: “No, absolutely not. I think those are two very different things.”
Anthropic prepares for a listing while OpenAI rules out 2026
Anthropic is preparing to enter the public markets, putting greater focus on how it would describe model safety, potential liability and risks to its business model. OpenAI CEO Sam Altman, by contrast, said on September 12 that OpenAI would not go public in 2026 because the company’s safety work was not yet complete.
Karp later expressed further doubts about OpenAI’s prospects as a public company, arguing that the scale of potential liabilities facing frontier AI could be difficult for traditional public markets to absorb. That remains his assessment of the industry’s liability structure and the capacity of capital markets; it does not represent an announced listing plan or cancellation by OpenAI.
Palantir sells software to governments and large enterprises, placing it in a related enterprise AI market alongside Anthropic and OpenAI. Karp’s comments therefore also reflect a clear industry position. For markets, the more immediate questions are whether OpenAI and Anthropic complete the model-testing agreement, whether Anthropic files an S-1, and how each company discloses model-safety incidents and related liability.