OpenAI, Anthropic, and Google DeepMind confirmed they have coordinated on AI safety measures for several weeks The three labs jointly unveiled cyber-focused AI models, safeguards, and access
OpenAI, Anthropic, and Google DeepMind confirmed this week that they have spent several weeks coordinating on AI safety measures, jointly unveiling a set of cyber-focused AI models, safeguards, and access programs developed collaboratively across the three organizations. The disclosure marks a notable departure from the labs’ typical public posture, which has generally emphasized competitive differentiation over shared initiatives, given that all three compete directly for enterprise customers, research talent, and compute infrastructure.
The specific focus on cybersecurity appears deliberate. As AI models have grown more capable at writing and analyzing code, they have simultaneously become useful for both defending against cyberattacks and, in the wrong hands, launching them, a dual-use problem that has drawn increasing attention from security researchers and policymakers alike. Coordinating on shared safeguards for this particular capability lets the three labs address a risk that affects all of them roughly equally, without requiring the kind of broader cooperation on model architecture or training methods that would touch more directly on their competitive positioning against each other.
The announcement lands in the same week as two related developments: a royal-hosted AI safety summit in Scotland bringing together executives from these same labs alongside Nvidia, and a separate reported incident in which researchers used Anthropic’s own Claude model to breach OpenAI’s internal systems as part of an authorized security exercise. Taken together, the three events paint a picture of an industry increasingly treating certain safety and security risks as shared problems requiring joint responses, even as the same companies continue to compete aggressively on nearly every other front, including model performance, pricing, and enterprise partnerships.
Details of exactly how the joint safeguards will be maintained going forward, including whether the three labs plan to continue coordinating on future safety releases or whether this represents a one-time collaborative effort tied specifically to the cybersecurity theme, have not been fully disclosed. Industry observers will likely watch whether this kind of cooperation becomes a recurring pattern or remains an isolated response to a specific, shared concern about AI-assisted cyberattacks.
The three labs’ joint focus on cybersecurity specifically, rather than a broader set of safety concerns such as model bias or misuse in generating disinformation, suggests the companies see cyber risk as an area where coordinated action is both urgent and less likely to require compromising competitively sensitive details about their own model architectures. Publishing shared safeguards for cyber-focused capabilities lets each lab demonstrate responsible behavior to regulators and the public without disclosing proprietary training techniques, a distinction that likely made this particular form of cooperation easier to agree on than deeper collaboration touching core model development would have been.
This post first appeared in OpenAI, Anthropic, and Google DeepMind Confirm Weeks of Safety Coordination
