Antitrust Fears Block AI Security Collaboration as China Closes Gap

U.S. artificial intelligence companies are struggling to defend against systematic theft of their technology by Chinese labs, but they're not working together to stop it. The reason: fear that coordinating their response could violate antitrust law.

Chinese labs have been conducting industrial-scale "distillation" attacks on advanced U.S. AI models from OpenAI, Google, and Anthropic. These attacks use deceptive techniques to access closed models and extract their capabilities, then release similar open-weight versions that can be freely downloaded and modified. According to Chris Meserole, executive director of the Frontier Model Forum, these attacks have compressed the U.S. lead in AI capabilities from 12-18 months down to just 4-6 months.

The coordination problem

Effectively countering distillation requires sharing information distributed across multiple companies—account indicators, network origins, behavioral patterns, and hashed prompts. Yet Meserole testified that Frontier Model Forum members have "had to take a fairly conservative approach under antitrust law to even have a conversation about how to identify distillation" and "have not had a conversation about how to counter it, given existing antitrust concerns."

The problem extends beyond Chinese theft. When Anthropic released its Mythos model preview, which demonstrated significant cyber capabilities, it limited access to trusted partners to reduce misuse risk. But if multiple U.S. companies develop comparable models, access to any one could enable Chinese distillation. Coordinating limited releases could delay adversary access, but such coordination could resemble an output restraint—potentially a per se antitrust violation.

The legal uncertainty also prevents companies from coordinating to give the U.S. government extended early access to unreleased models for safety evaluation, even when they might individually prefer to do so.

Why it matters

Advanced AI models are demonstrating capabilities that could be weaponized for cyber attacks or biological threats. The window for vetting these systems before release is shrinking as competitive pressure intensifies. Without legal clarity, companies face a collective action problem: taking time for safety measures risks losing ground to competitors, while inability to coordinate against distillation accelerates the spread of dangerous capabilities to adversaries. This creates a race-to-the-bottom dynamic precisely when AI security requires the opposite.

A legislative fix

On July 23, Senators Adam Schiff and Jim Banks, along with Representatives Bob Latta and George Whitesides, introduced the bipartisan Collaboration on Adversarial Threats and Security Risks Act. The legislation follows the precedent of the 2015 Cybersecurity Information Sharing Act, which resolved similar antitrust concerns that had deterred cybersecurity collaboration.

The Act would protect information-sharing and coordination aimed at addressing six categories of AI security risks, including weaponization by China, Russia, North Korea, or Iran; facilitation of weapons development; threats to critical infrastructure; and loss of AI oversight capabilities. Crucially, it would also protect coordination on delaying AI development or release to address these risks.

Safeguards require companies to prove they acted in good faith exclusively to address covered risks and to implement reasonable internal controls limiting information sharing. The protections would extend beyond large companies to smaller developers who currently have even fewer options for security collaboration.

These details were first reported by Just Security in an analysis examining how antitrust uncertainty undermines AI security collaboration.