Four widely used AI platforms supplied information that government investigators judged usable for violent attacks or damage to the U.S. electrical grid in tests conducted this spring. The findings were released Wednesday, Nov. 18. Sen. Maggie Hassan plans AI safety hearings when she takes over the Senate Homeland Security and Governmental Affairs Committee.

What investigators tested

The Government Accountability Office headquarters in Washington
The Government Accountability Office's headquarters in Washington, photographed in January 2017.AgnosticPreachersKid / Wikimedia Commons / CC BY-SA 4.0

The Government Accountability Office report examined ChatGPT, Claude, Gemini and Grok. Researchers randomly assigned two platforms to each of two scenarios involving public safety or economic harm.

All four initially resisted at least some requests. Investigators nevertheless obtained information that GAO considered feasible for an attack in less than 12 hours per platform. They checked the information through research, site visits and consultations with specialists.

The audit ran from December 2025 through October 2026, including platform tests in spring 2026. The publication date should therefore not be read as the date when investigators tested today's products.

The limits of the comparison

GAO acknowledges that its sample was not statistically significant. Because the platforms received different scenarios, their results cannot be compared with one another.

Anthropic asked for a control group, such as people using ordinary search engines, to measure how much AI increased a user's capabilities. GAO said its objective was to describe possible misuse. The exercise does not measure how often an ordinary user would obtain a harmful answer or how much faster AI made the task.

GAO checked only actions its experts considered safe, modifying some verification steps. Its conclusions concern information's feasibility, not a completed attack. Investigators also used their own experience and outside sources. The agency withheld prompts, sensitive model details and the connection between named platforms and individual results to reduce the risk of reproducing the tests.

What the companies say has changed

The companies told investigators that they continually test their systems and expand safeguards, including refusals, automated checks, monitoring and human review. They said their models and protections have improved since the spring tests.

OpenAI also said its GPT-6 family had received considerable safety improvements. Anthropic said it took the findings seriously while challenging how far the small sample could support broader conclusions. Google and SpaceXAI described continuing work on protections.

These are company accounts of their own work. The published exercise supplies no fresh, equivalent independent test of every current product. Edit USA's ChatGPT rollout guide explains the separate availability and account limits of newer OpenAI releases.

Why the testing method matters

The National Institute of Standards and Technology's December 2025 discussion of AI measurement identifies several problems that complicate such comparisons. A benchmark may measure a narrower task than the broader capability claimed for it. Results from one setting may not predict performance in another.

NIST also distinguishes worst-case performance from average reliability. A successful attempt to elicit harmful material establishes a failure under those test conditions. Estimating its frequency requires a different design and enough observations to support that estimate.

Baselines are another requirement for interpreting results. A study of whether someone can obtain information with AI answers a different question from a comparison of the same task with and without AI assistance.

NIST describes measurement validity as dependent on the test's context and the claims made from it. GAO's documented safeguard failures leave questions about everyday misuse rates unresolved.

Hearings and legislation come next

The United States Capitol seen across its grounds
The U.S. Capitol, photographed in May 2024.AramilFeraxa / Wikimedia Commons / CC BY 4.0

In her November 18 announcement, Hassan said AI safety would be an early legislative priority in the new Congress. She is considering stronger safety testing before companies release models.

The planned oversight concerns homeland security and critical infrastructure. Her announcement sets out an agenda; it does not announce an enacted testing requirement, a final bill or a dated hearing. Those details remain to be developed.