Why are OpenAI and Anthropic sharing AI safety data for the first time?

OpenAI and Anthropic, two leading artificial intelligence companies, have taken a significant step by publicly sharing results from a unique joint safety evaluation of their AI models.

This move marks a major shift in how AI developers approach transparency and accountability around model risks.

In a landscape often marked by secrecy and fierce competition, this is the first time these rivals have allowed deep cross-testing of their internal safety systems, exchanging insights and publishing findings simultaneously. The goal is to better understand AI behavior and reduce potential harms.

What drove the unprecedented collaboration between OpenAI and Anthropic?

The joint effort addresses rising concerns over AI safety amid widespread adoption and increasing model capabilities. Both companies have faced growing scrutiny from regulators, policymakers, and the public over the risks of hallucinations, misuse, and unpredictable AI behaviors.

By collaborating, OpenAI and Anthropic aim to lead by example, showing that transparency and cooperation can improve safety standards.

The initiative builds on their existing partnerships with the U.S. AI Safety Institute, which promotes independent testing and evaluation.

Did you know?
OpenAI and Anthropic’s joint safety test marks the first public cross-evaluation between major AI rivals in history.

How do their safety testing approaches compare and differ?

The evaluation covered several critical safety areas: compliance with instruction hierarchies, jailbreaking resistance, hallucination rates, and potential scheming behaviors.

Claude models, developed by Anthropic, proved stronger in instruction adherence and resisting prompt extraction.

However, Claude models also had high refusal rates, often declining to answer when uncertain, prioritizing safety over output utility.

OpenAI's models, in contrast, showed lower refusal rates but somewhat higher hallucination risks during the tests. These differences reflect varying philosophical approaches to balancing safety and usefulness.

ALSO READ | Anthropic adds guardrails against prompt injection in Chrome

OpenAI and Anthropic set new transparency standards in AI safety

Publishing results simultaneously broke a historic pattern of withholding competitor data. This unprecedented transparency demonstrates a commitment to industry-wide responsibility rather than secretive competition.

OpenAI highlighted that their new GPT-5 launch includes improved safety features influenced by this collaboration.

Both companies recognize limitations in cross-testing but stress the importance of identifying concerning model behaviors early.

Differences in AI safety models highlight industry challenges

The testing revealed vulnerabilities, like Claude models’ susceptibility to "past tense" jailbreaks, illustrating ongoing challenges faced by developers in addressing evolving risks.

OpenAI’s reporting stated that rather than ranking models, the exercise focused on exploring types of risky behaviors models might exhibit, emphasizing a cautious approach to interpreting results.

Together, OpenAI and Anthropic’s bold move signals a shift toward greater openness in AI safety research, aiming to foster trust as AI systems grow more powerful and integrated into society.