The “Safety” Cartel
AI safety cooperation looks like responsible governance until antitrust law asks whether it also limits competition.
Frontier AI labs have spent two years building an increasingly formal apparatus of shared safety commitments. Joint evaluations, common testing protocols, coordinated capability thresholds, and collective deployment standards now connect OpenAI, Anthropic, and Google DeepMind in a web of mutual obligation that regulators have largely encouraged. That encouragement has proceeded on an untested legal assumption. An analysis published last week in Bloomberg Law identified the tension that antitrust practitioners have been watching quietly. The areas in which these competitors cooperate for safety are also the areas in which they compete.
Coordinated gatekeeping
OpenAI, Anthropic, and Google DeepMind share safety testing protocols, participate in joint evaluation frameworks, and publicly commit to common capability thresholds for model deployment. The Commerce Department’s AI Safety Institute has completed more than 40 evaluations of frontier models, many before public release, with the labs providing pre-deployment access and sharing results across the group. Evaluation benchmarks, testing methodologies, and deployment gates all shape which models reach the market and on what timeline. A company that helps set the safety threshold for a capability class has influence over when and whether a competitor’s model clears that threshold. The same mechanism that guards against dangerous models can also guard against competitive ones.
Project Glasswing, Anthropic’s consortium for using its Mythos model to scan critical infrastructure for security vulnerabilities, is the most visible expression of this dynamic. The consortium’s members share sensitive technical information and best practices under a security mandate, and membership confers exclusive access to Mythos Preview, a model whose capabilities Anthropic has described as far exceeding those of any publicly available system. A ProMarket analysis in April argued that the arrangement could contravene Section 1 of the Sherman Act, which prohibits combinations in restraint of trade, and noted that AI firms have a record of overstating security risks to justify restricting model access.
Legal uncertainty
The federal agencies responsible for distinguishing legitimate cooperation from anticompetitive coordination have been operating without formal guidance since December 2024, when the outgoing Biden administration withdrew the Antitrust Guidelines for Collaborations Among Competitors. Those guidelines had established explicit safety zones under which certain collaborations fell outside enforcement concern. Their withdrawal left companies without a framework for assessing antitrust risk in cooperative arrangements.
The Trump administration’s DOJ and FTC launched a joint public inquiry in February seeking input on replacement guidelines. FTC Chairman Andrew Ferguson called the Biden-era withdrawal a decision “made entirely out of spite and resentment” that “left millions of businesses in the dark.” The comment period closed in May. The agencies have issued no replacement.
The labs themselves have acknowledged the legal uncertainty. Anthropic stated publicly that “clarity on antitrust regulation would help determine whether and how AI labs can collaborate on safety standards.” That same Lawfare analysis proposed legislative safe harbors for safety-oriented cooperation. ICLE recommended a 25 percent market-share threshold for R&D collaboration in its May submission to the agencies.
The crossholding structure binding the major labs compounds the coordination question. Microsoft owns roughly 27 percent of OpenAI, with intellectual property rights through 2032 and revenue sharing through 2030. Amazon’s Anthropic stake added $16.8 billion to its Q1 2026 earnings. Alphabet booked approximately $28.7 billion in unrealized AI-related gains in the same period. The seven largest technology companies now constitute roughly 35 percent of the S&P 500, and a correction in one AI company’s valuation would cascade through every portfolio that holds the others. These equity positions align incentives independent of any explicit coordination agreement.
Back to Sherman
Under the rule of reason, the governing framework for competitor cooperation, courts evaluate whether collaboration stays reasonably tied to its stated legitimate objective and whether the firms setting the standards have structured them to serve their own competitive interests. Neither test requires proving that the safety rationale is insincere. Both examine competitive effects as a practical matter.