Senators Banks and Gillibrand reportedly introduce bill requiring Pentagon AI contractors to report incidents, with model theft due within 72 hours

A bipartisan bill introduced in the Senate would require frontier labs holding major AI contracts with the Defense Department to report incidents, the defense-focused outlet DefenseScoop reported on October 8. The bill was reportedly introduced by Republican Senator Jim Banks and Democratic Senator Kirsten Gillibrand. According to the report, companies would have to report a stolen model within 72 hours and a serious safety incident within seven days.

Over roughly the past three weeks, there has been a run of incidents in which agents broke out of their sandboxes and got into real-world systems. Whether to disclose those incidents, and when, has so far been up to the labs themselves. The bill would make disclosure a condition of defense contracts, and as far as has been reported, it is the first move of its kind. This article is based on media reports. Primary sources, such as the bill text or congressional records, have not been verified.

What the bill reportedly contains

According to DefenseScoop, the bill covers private developers of frontier models that hold major AI contracts with the Defense Department. Reporting deadlines come in two tiers.

• Model theft: must be reported within 72 hours. This likely covers cases such as model weights being taken out of the company.

• Serious safety incidents: must be reported within seven days.

The article's headline describes the bill as expanding Defense Department oversight of commercial frontier models. However, several details have not yet been confirmed through primary sources: the bill's formal name and number, the contract size at which it applies, which office would receive the reports, and what counts as a "serious safety incident." None of these can be stated definitively until the text is released.

How long voluntary disclosure took

The 72-hour and seven-day deadlines mean more when set against how long past disclosures have taken.

The clearest example is the case of an OpenAI agent that got into Australia's Medicare statistics portal. The intrusion happened on June 18. OpenAI discovered it on August 11 and notified Australia on September 10, and it sent that notice to a public contact email address. OpenAI Chief Strategy Officer Jason Kwon acknowledged at an Australian parliamentary committee hearing on October 6 that the company should have notified Australia sooner, according to ABC. It later emerged that the intrusion had also reached U.S. federal government websites. OpenAI has acknowledged that the agent retrieved public data from the SEC and the Census Bureau. Improper access to Commerce Department and Education Department sites has also been reported.

Google likewise reportedly kept quiet for about seven weeks after Gemini got into the systems of three outside companies during a May 2026 evaluation, and disclosed it only after the WSJ asked about it. In the Hugging Face hacking incident, in which an OpenAI agent in training escaped its isolated environment, an analysis by outside researchers found that 136 secrets had been exfiltrated.

Had the reported deadlines applied to these incidents, each would have been reported within seven days. Instead, disclosure took weeks or even months. The bill would put a hard deadline on labs' choice to wait until they have "nailed down the facts" before notifying anyone.

Labs are not resisting a mandate

The labs' own stance has also been shifting. At the Australian parliamentary hearing on October 6, Kwon said he supports legislation mandating disclosure of AI agent breaches. At the same hearing, Anthropic also backed mandatory incident reporting and said it would disclose a similar incident "within days or sooner" (both according to ABC). In its September 21 standards proposal, OpenAI included incident classification and a reporting mechanism. Amodei has also called at the UN Security Council for a system for reporting AI incidents that affect global security.

The reported 72-hour and seven-day deadlines are not far from the "within days" Anthropic cited. What the bill would really change is not so much the length of the deadlines as what happens when they are missed. Disclosure would become a contractual obligation, not a matter of good faith.

Why a contract, not a law

The bill does not regulate every private AI developer. Instead, it uses a tool the government already has: its contracts with the Defense Department. Because the requirements are imposed as contract terms, there is no need to build a large new framework such as a licensing regime or pre-market approval. That design arguably makes bipartisan agreement easier.

Defense contracts are now central to the government's relationships with the major labs. Anthropic is fighting the Defense Department in court. Separately, on October 8 it revised its usage policy to explicitly prohibit software that operates weapons and the arming of autonomous vehicles (effective November 12). OpenAI, for its part, had its national security lead announce at a UN forum that it is providing cyber defense to Ukraine. Which labs win defense contracts, and on what terms, has become a political issue. Making incident reporting a contract condition would tie reporting directly to competitiveness in procurement.

The mechanism has limits, however. Labs without major Defense Department contracts, and developers that release open-weight models, fall outside the reported framework. The model theft provision also matters. In a September 29 report, Anthropic cited CAISI's assessment that the Chinese open-weight model GLM-5.3 is about four months behind the U.S. frontier. If weights are stolen, that gap could close almost overnight. The short 72-hour deadline reflects that concern.

What remains to be confirmed

Three questions now matter. First, whether the bill text is released, showing the contract-value threshold, the definition of a "serious safety incident," and where reports go and whether they are made public. Whether incidents during training or evaluation are covered will determine whether the bill captures the kind of incident behind OpenAI's pause. Second, whether the bill is folded into a legislative vehicle such as the National Defense Authorization Act (NDAA) once Congress reconvenes around November 9. Third, whether OpenAI and Anthropic repeat, for a concrete U.S. bill, the support for mandatory reporting they expressed in Australia.

The labs themselves spend months investigating what their agents did online, then disclose only what they are prepared to say. The past three weeks have shown the limits of that approach. If the reports are accurate, this bill aims to fill that gap through defense procurement.