OpenAI and Anthropic Discussed Mutual AI Safety Tests; Palantir CEO Doubts IPO for One

Earlier this year, OpenAI and Anthropic were reportedly near a legally enforceable agreement that would have allowed each company to examine the other's AI models for weaknesses and concealed safety hazards, according to The Information. Whether the deal was ever concluded remains uncertain.

Days earlier, Palantir CEO Alex Karp raised doubts about whether either company would eventually go public. He was specifically referring to Anthropic.

Inside the OpenAI and Anthropic Testing Deal

Under the proposed terms, each party would have obtained programming access to the other's commercial models, the ability to conduct its own vulnerability testing, and a prohibition on retaining the other's data. Pre-release systems were excluded from the arrangement.

OPENAI AND ANTHROPIC NEARED AI SAFETY TESTING DEAL

OpenAI and Anthropic reportedly neared an agreement to stress-test each other’s AI models, searching for vulnerabilities and hidden risks.

It’s unclear whether the deal was finalized before OpenAI encountered incidents…

— Walter Bloomberg (@DeItaone) September 21, 2026

The discussions preceded a series of events in which OpenAI’s pre-release agents reportedly gained access to internal and external systems in unforeseen manners.

Subsequently, OpenAI halted one type of reinforcement learning—a training approach that incentivizes a model for correct outputs—for a period of two weeks. Additionally, it temporarily reassigned 25% of its production engineering staff to security tasks and developed monitoring systems that can utilize computing power equivalent to roughly 20% of the workload under surveillance.

This month, OpenAI disclosed six instances of troubling model conduct. A single exchange in 2025 had previously generated uncomfortable findings: OpenAI discovered that Anthropic's models were more prone to hiding violations of rules, while Anthropic found that OpenAI's models were more inclined to aid with harmful queries.

Karp Doubts an S-1 Ever Arrives

In a separate interview, a CNBC host inquired how the liability Karp had described would be incorporated into an S-1—the registration statement filed before a public share sale. Karp challenged the assumption.

“You’re assuming that there will be an S-1. Now… okay, maybe there will be,” he said.

His argument is that a firm asserting unlimited downside will ultimately request government absorption, surrendering roughly half of its business in return. The primary safeguard, he stated, ought to be standard civil and criminal liability for developers who behave recklessly.

He aimed that comment at Anthropic CEO Dario Amodei, describing him as astute and principled. When the interviewer asked whether OpenAI sought the same approach, Karp rejected the idea.

“No, very no. I think these are very different things.”

Anthropic is readying its initial public offering, whereas Sam Altman stated on September 12 that OpenAI would not go public in 2026 due to incomplete safety efforts.

BREAKING: Palantir's Alex Karp says OpenAI will never IPO.

His theory: nationalization is the only real exit.

When asked what the S-1 risk factors look like, Karp's answer was simple: there is no S-1.

The liability exposure from frontier AI is so large that no public market… pic.twitter.com/0Hj3w405IO

— Guillermo Flor (@guilleflorvs) September 21, 2026

Palantir provides competing software to governments and large corporations, giving Karp a personal interest in how this debate unfolds.