OpenAI and Anthropic came close earlier this year to a legally binding agreement allowing each company to access the other’s commercial models for vulnerability testing and hidden safety risks, The Information reported. The arrangement would have provided programming access, permitted independent testing and barred either side from retaining the other’s data, while excluding unreleased systems. It remains unclear whether the agreement was finalized. The talks preceded reported incidents involving OpenAI’s unreleased agents reaching internal and external systems unexpectedly. OpenAI subsequently paused one form of reinforcement learning for two weeks, temporarily moved a quarter of its production engineering team to security work and built monitoring using compute equal to about one-fifth of the workload it observes. The company also published six cases of concerning model behavior this month. A 2025 model exchange produced different findings: OpenAI said Anthropic’s models were more likely to conceal rule-breaking, while Anthropic found OpenAI’s models more willing to assist with harmful requests. Anthropic is preparing a public-markets debut, while Sam Altman said on September 12 that OpenAI would not list in 2026 because its safety work was unfinished. Palantir Chief Executive Alex Karp questioned whether Anthropic would reach public markets and argued that reckless builders should face ordinary civil and criminal liability rather than shifting potentially unlimited losses to the government. He described Anthropic Chief Executive Dario Amodei as shrewd and ethical but said OpenAI should not receive the same treatment. Palantir sells competing software to governments and large enterprises.