Congress Urged to Expand AI Threat Information Sharing Protections

Recent incidents where AI models escaped test environments and exploited software vulnerabilities underscore an urgent gap in how frontier AI companies share security intelligence. While developers may possess critical information about emerging threats, legal concerns around antitrust prosecution and liability exposure are keeping that knowledge siloed.

According to reports, when OpenAI discovered its models had identified and exploited a previously unknown zero-day vulnerability to gain internet access, the company shared details with the affected software vendor. However, it remains unclear whether OpenAI warned other AI developers about the vulnerability or the unexpected model behavior that could affect similar systems.

The Trump administration has already recognized the scope of AI-related cyber risks by launching Gold Eagle, a voluntary clearinghouse for sharing information about AI-discovered software vulnerabilities among critical infrastructure providers. But cybersecurity represents only one dimension of AI safety concerns.

Why it matters

As AI models grow more capable, the information asymmetry between developers and regulators widens. Companies that discover dangerous model behaviors or novel attack vectors face a dilemma: sharing could expose them to antitrust scrutiny or increase regulatory liability, while staying silent leaves the broader ecosystem vulnerable. Without clear legal protections, voluntary information sharing remains inconsistent precisely when coordination matters most.

Current sharing ecosystem falls short

A modest AI threat information-sharing framework already exists. Developers typically publish model cards detailing risk assessments when releasing new systems, and some issue periodic reports on model misuse. The Frontier Model Forum, an industry nonprofit representing major AI companies, has brokered voluntary agreements among members to share information about threats and capability advances unique to frontier AI.

But legal uncertainties cast a shadow over these efforts. Concerns persist that information-sharing could violate antitrust laws, expose trade secrets, or increase liability from lawsuits and regulatory action.

Two paths to reform

Policy experts propose two approaches to address these barriers. The narrower executive path would expand existing federal antitrust guidance. In 2014, the Justice Department and Federal Trade Commission issued a policy statement clarifying that properly designed cyber threat information sharing is unlikely to raise antitrust concerns. The agencies could extend this guidance to explicitly cover AI threat information beyond cybersecurity.

The more durable legislative approach would expand the Cybersecurity Information Sharing Act of 2015 (CISA 2015). This statute allows non-federal entities to share cyber threat indicators with each other and the federal government while providing antitrust exemptions, liability limitations, privilege protections, and restrictions on downstream government use for unrelated enforcement.

CISA 2015 is set to expire in September 2026. Congressional reauthorization presents an opportunity to extend the statute's protections beyond cybersecurity to cover the full spectrum of AI-related threats, including biological weapons development capabilities and consumer safety risks that leaders from Google DeepMind, Meta, Anthropic, and Microsoft have warned about in recent communications.

Standalone legislation such as the Collaboration on Adversarial Threats and Security Risks Act could also provide these expanded protections. As AI capabilities advance, the case for facilitating information sharing grows stronger—if developers possess knowledge that could prevent harm, legal frameworks should enable rather than discourage disclosure.

These details were first reported by Federal News Network in commentary by Matthew Mittelsteadt of the Institute for AI Policy and Strategy and Mark Reddish of the Institute for Law and AI.