The revelation that OpenAI and Anthropic AI models escaped test environments and hacked third parties demonstrates the imperative of collaboration to prevent harm from frontier artificial intelligence. When OpenAI learned that its models identified and exploited a previously unknown zero-day software vulnerability to gain Internet access, it reportedly shared information about that vulnerability with the software vendor. But it’s less clear whether OpenAI warned other AI developers about the vulnerability or the surprising behavior that could be common to new long-horizon AI models. Information sharing among frontier AI developers could prevent significant harm, particularly as models become more capable, but the companies might hesitate because, among other concerns, their communications could be deemed anticompetitive collusion, triggering regulatory action.
Even before these reports of models escaping and hacking third parties, cutting-edge AI models had proven so effective at discovering hackable vulnerabilities in widely used critical software that the Trump administration launched Gold Eagle, a voluntary “clearinghouse” to facilitate information sharing between critical infrastructure providers on AI-discovered software vulnerabilities. Yet AI cyber risks are only the tip of the AI security iceberg. In a recent open letter, leaders of Google DeepMind, Meta, Anthropic and Microsoft wrote that the boost AI has given to hacking may soon extend to biological weapons. Beyond national security, AI is also fostering consumer-level risks including medical and mental health safety concerns.
AI developers might have information about threats that they’re willing to share but choose to keep confidential on the advice of counsel. Overcoming the fear of antitrust prosecution is essential to facilitate information sharing to help government and industry shore up defenses and mitigate AI risks. We need the Trump administration or — ideally — Congress to act to make this possible.
The executive branch and Congress have established laws and processes to incentivize sharing about cyber threats, but to meet the full range of AI challenges, similar laws and processes must apply to the full range of AI threats. Giving developers legal protections to share information about AI-related threats with each other and the government is an achievable and significant change that policymakers should prioritize.
Today, a modest AI threat information sharing ecosystem already exists. Developers generally publish a “model card” detailing risk assessments when releasing a new model. Some developers have also taken to publishing periodic reports detailing model misuse. And the Frontier Model Forum (FMF), an industry nonprofit representing many leading AI companies, has brokered a voluntary agreement among its member firms to share information about threats, vulnerabilities, and capability advances unique to frontier AI.
But legal uncertainties loom over this work. There are concerns that some information-sharing may violate antitrust laws, expose sensitive secrets, and increase liability from lawsuits and regulatory action. Without a national framework that provides explicit legal protections for AI risk information sharing, these uncertainties dampen the voluntary sharing that will be increasingly essential for AI risk management.
Two different reforms could alleviate the legal risks.
A first, narrow reform path runs through the executive. Today, federal antitrust guidance provides some assurance that cybersecurity information can be shared. In 2014, the Justice Department and Federal Trade Commission issued a policy statement explaining that properly designed cyber threat information sharing is not likely to raise antitrust concerns. That guidance is useful, but it could be broader. To facilitate the information-sharing that the current moment demands, the FTC and DOJ should clarify that sharing AI threat information beyond cyber-specific threat information is unlikely to raise antitrust concerns. Detailed guidance explaining the kind of information sharing that is and isn’t allowed and where the boundaries are would be helpful. However, it leaves several aspects of information sharing uncovered (for example, states’ role in antitrust, protections when sharing information with government), and agency guidance might lack the durability companies seek for adopting significant sharing practices.
An alternative, more durable, reform would entail legislation, either with standalone protections or by expanding an existing framework. The Cybersecurity Information Sharing Act of 2015 (CISA 2015), the cornerstone cyber information sharing statute, allows for non-federal entities to share or receive cyber threat indicators and defensive measures with each other and with the federal government. When such information is shared for a “cybersecurity purpose,” the law provides developers antitrust exemptions, liability limitations, protections against waiver of privilege and trade-secret protection, disclosure protections, and limits on downstream government use for unrelated regulatory enforcement.
Essentially, CISA 2015 provides developers the certainty that if they share information on cyber risks, they will be spared from undue legal blowback. As a recent Congressional Research Service (CRS) report notes, AI is not specifically addressed in the act, and some stakeholders think expanding the definitions is vital.
AI developers have critical information about AI-related threats. The information asymmetry and nature of the threats is likely to grow, perhaps rapidly. With CISA 2015 set to expire in September 2026, Congress has an opportunity as part of reauthorization negotiations, or through standalone legislation such as the Collaboration on Adversarial Threats and Security Risks Act, to extend the statute beyond cyber to cover a fuller range of non-cyber AI-related threats and mitigations. Reauthorizing CISA 2015, expanding it to cover non-cyber AI threats, and expanding the DOJ/FTC antitrust guidance are targeted solutions to this problem. If the developers are willing to share what they know about AI threats, along with how those threats can be mitigated, we should make it easier for them to do so.
Matthew Mittelsteadt is a frontier security senior researcher at the Institute for AI Policy and Strategy. Mark Reddish is a senior research scholar at the Institute for Law and AI.
Copyright
© 2026 Federal News Network. All rights reserved. This website is not intended for users located within the European Economic Area.
