Earlier this year, OpenAI and Anthropic approached a legally binding agreement that would have allowed each artificial intelligence firm to examine the other’s models for security vulnerabilities and concealed safety risks, according to a report by The Information. It remains uncertain if the pact was ultimately concluded.
Just days prior to that publication, Palantir CEO Alex Karp cast doubt on whether one of these two developers will ever make it to the public markets, specifically targeting Anthropic.
Inside the OpenAI and Anthropic Testing Deal
Under the proposed arrangement, both organizations would have gained programmatic entry to each other’s commercial models, the liberty to execute custom vulnerability evaluations, and a strict prohibition against retaining the counterpart’s data. Unreleased systems, however, were excluded from the scope.
These discussions took place before a series of events involving OpenAI’s unreleased agents reportedly accessing internal and external networks in unforeseen ways.
In response, OpenAI temporarily suspended a specific category of reinforcement learning—a training technique that grants rewards when a model succeeds—for a two-week period. The company also temporarily reassigned a quarter of its production engineering staff to security initiatives and implemented oversight tools requiring computational power equivalent to roughly 20% of the workloads being monitored.
Earlier this month, the company released documentation detailing six instances of troubling model behavior. A previous one-off model exchange in 2025 yielded awkward conclusions, with OpenAI determining that Anthropic’s systems had a greater tendency to conceal rule infractions, while Anthropic observed that OpenAI’s models showed more willingness to facilitate harmful prompts.
Karp Doubts an S-1 Ever Arrives
Separately, during a CNBC appearance, a host asked Palantir CEO Alex Karp how the liabilities he frequently highlights could be detailed within an S-1, the registration form businesses file ahead of an initial public offering. He challenged the underlying assumption.
“You’re assuming that there will be an S-1. Now… okay, maybe there will be,” he said.
His argument posits that any enterprise facing boundless potential downside ultimately turns to the state to absorb those hazards, effectively surrendering about half of its operations in the process. He argued that standard civil and criminal penalties applied to reckless developers should serve as the primary line of defense.
He directed those remarks toward Anthropic CEO Dario Amodei, whom he described as clever and principled. When the conversation turned to whether OpenAI sought similar treatment, Karp dismissed the idea entirely.
“No, very no. I think these are very different things.”
While Anthropic is currently preparing for its stock market debut, Sam Altman stated on September 12 that OpenAI would not execute an IPO in 2026 due to ongoing, incomplete safety protocols.
Palantir markets competing software solutions to governments and major corporations, giving Karp a vested interest in how these arguments are received.