OpenAI and Anthropic Consider Mutual AI Tests; Palantir on IPO

OpenAI and Anthropic held talks about running mutual tests of their AI systems to compare safety and behavior across models. Company officials and people involved in safety work described the conversations as part of broader industry exchanges among developers, researchers and funders focused on how large models handle risky or unexpected inputs.

Those involved said mutual testing would involve each lab evaluating the other’s models on agreed scenarios to find failure modes, safety blind spots and areas for improvement. Proposed methods include shared benchmarks, adversarial prompts, tracing decision paths and coordinated red-team exercises that simulate misuse. The discussions did not result in a formal program, and the firms did not disclose a schedule or which teams would take part.

People familiar with the talks noted legal and competitive issues that would need to be resolved before any joint testing. Topics mentioned include intellectual property protection, nondisclosure of proprietary model internals and narrow rules on what findings could be shared or published. Both companies have released some research on safety practices while keeping other technical details private as they commercialize products.

Palantir chief executive Alex Karp raised the possibility that one of the leading AI labs could remain privately held indefinitely. He did not identify a firm or offer a timetable. The remark highlighted questions among investors and policymakers about how governance, funding and commercial agreements will affect long-term paths for major AI developers.

OpenAI and Anthropic remain private entities that have taken large private investments and struck commercial partnerships to fund model development and cloud infrastructure. The talks on cross-testing reflect ongoing industry attention to methods for validating model behavior while protecting proprietary technology.