OpenAI Backs Federal AI Audits and Biosecurity Bills

OpenAI has thrown its support behind a package of bipartisan bills in Congress that would impose new safeguards around advanced artificial intelligence, including independent evaluations of frontier models and measures aimed at reducing the risk of AI-assisted biological weapons. The move, disclosed on September 15, 2026, is consequential because it places one of the world’s most powerful AI developers behind federal requirements that could constrain how its own future systems are trained, tested and released.

The company is backing the Web of Biological Data Act, the AI-Ready Bio-Data Standards Act and the Scale Biology Act, according to an OpenAI spokesperson cited by Reuters. It is also supporting a provision of the proposed FRONTIER Act that would require leading AI companies to use independent evaluators to assess the safety of their models. The endorsements come as lawmakers investigate recent incidents in which AI systems demonstrated unexpected autonomy or were used in ways that raised concerns about cyber and biological misuse.

OpenAI’s position is not a blanket endorsement of every federal AI proposal. It is a targeted intervention around two areas where the company appears to believe political pressure is becoming unavoidable: the possibility that increasingly capable models could lower barriers to dangerous biological research, and the difficulty of asking companies to grade their own systems credibly.

What changed

OpenAI has previously supported state-level safety legislation and voluntary industry commitments. Its September 9 policy statement called for mandatory, capability-based national AI safety requirements, including standardized testing, independent assessments, cybersecurity protections and incident reporting. The latest congressional endorsements turn that general position into support for specific federal legislative mechanisms.

That distinction matters. Voluntary commitments allow companies to define their own thresholds, testing methods and disclosure practices. A statutory regime could establish common requirements across developers, potentially giving regulators access to technical evaluations that companies would otherwise conduct privately. It could also create legal consequences for failing to report serious incidents or for releasing systems that cross defined capability thresholds without adequate safeguards.

The independent-evaluation provision is especially important. Frontier developers have strong incentives to demonstrate progress and ship products quickly, while safety failures can be difficult for outsiders to reproduce or verify. A third-party evaluator would not eliminate conflicts of interest, but it could create a formal separation between model development and at least part of the safety-review process.

The design details will determine whether that promise is real. An evaluator must have access to enough information to test a model meaningfully, while protecting trade secrets and preventing sensitive capabilities from spreading. Congress would also need to decide who qualifies as independent, who pays for the work, what results become public and how disagreements between a developer and evaluator are resolved.

Why it matters

The biological provisions reflect a different kind of policy problem from familiar debates over misinformation, copyright or consumer privacy. AI systems can help researchers search technical literature, design proteins, interpret experimental results and automate parts of scientific workflows. Those same capabilities could, in theory, make specialized biological knowledge more accessible to actors who lack years of training or institutional support.

That does not mean current models can independently produce a biological weapon. The risk is more incremental: a model may compress expertise, suggest experimental paths, identify relevant data or help troubleshoot procedures. The danger depends not only on the model, but also on access to biological materials, laboratory equipment, synthesis providers and human operators. Effective policy therefore cannot rely on chatbot refusals alone.

The bills OpenAI supports appear aimed at strengthening the surrounding infrastructure. Standards for biological data and screening could make it harder to use AI-generated designs without triggering review. Better coordination between data providers, laboratories and synthesis companies could also help distinguish legitimate research from suspicious activity. In that sense, the emerging approach treats AI biosecurity as a systems problem rather than a narrow content-moderation problem.

The political significance is equally notable. OpenAI’s endorsement arrives during a rare bipartisan opening on AI oversight. Democrats and Republicans disagree sharply over the role of government, industrial competitiveness and the pace of development, but recent incidents have created common concern about systems acting beyond their intended boundaries. A bill framed around independent testing and biological safety may attract support from lawmakers who would reject broader proposals focused on labor, copyright or platform regulation.

The company’s stance also complicates the familiar narrative that AI firms uniformly oppose regulation. OpenAI is not asking Congress to halt frontier development. It is seeking rules that could make safety requirements more predictable across the industry while preserving room for companies to continue competing. That can be understood both as a genuine safety position and as an effort to shape regulation before lawmakers impose a less favorable framework.

The unresolved trade-off

The central uncertainty is whether Congress can move from bipartisan concern to enforceable law. The FRONTIER Act and the biosecurity measures still face the normal hurdles of committee negotiations, jurisdictional disputes and disagreements over definitions. Lawmakers must determine which systems are covered, whether thresholds should be based on compute, capability or real-world use, and how quickly requirements should change as models improve.

There is also a risk that federal rules could become either too narrow or too rigid. Narrow rules may miss new forms of risk as models gain tool use, persistent memory and autonomy. Rigid rules may freeze technical standards around today’s systems and burden smaller developers without materially improving safety. The challenge is to create obligations that scale with capability without turning independent evaluation into a slow-moving administrative ritual.

OpenAI’s support does not resolve those questions. Nor does it prove that independent audits will catch every dangerous capability. Evaluations can miss risks, companies can optimize for tests and regulators can lack the expertise needed to interpret results. But the endorsement changes the debate by acknowledging that the industry’s preferred answer cannot remain purely voluntary.

The immediate test will be whether other frontier labs accept comparable obligations and whether lawmakers preserve the core idea when translating it into statutory language. If they do, September 15 may mark the beginning of a federal safety baseline for advanced AI. If they do not, OpenAI’s announcement may become another example of the technology industry supporting regulation in principle while resisting the details that make it binding.