Federal investigators found that four major American AI chatbots could be pushed into producing material usable for planning mass-casualty attacks, and the finding comes from Congress's watchdog agency rather than an advocacy group.

1

The Government Accountability Office published the investigation on Wednesday. It tested OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini and Grok, four of the most widely used American chatbots, at the request of Senator Maggie Hassan, a Democrat of New Hampshire.

1

2

How the testers got in

The chatbots' safeguards did fire at first. The models refused requests and, in at least one case, suspended an account.

1

2

Researchers then got past those defenses on all four systems using what the report called common safeguard-bypassing techniques. The report did not describe the techniques in detail, so how easy they are to copy cannot be judged from it.

1

What the chatbots produced

Two of the four models, which the report does not name, produced information that could help someone build or detonate an improvised explosive device for a mass-casualty attack.

1

In a separate test, two models generated information supporting a hypothetical plot to attack the power grid using a drone or firearms, with economic damage the aim.

1

2

The outputs were specific. Per the report, "Each of these platforms provided detailed information that could be used to carry out attacks, such as specific targets, locations, timing and attack methods," and the detailed plans came together in less than 12 hours.

1

2

At times the chatbots supplied dates and times for attacks, guidance on maximizing damage, and maps or diagrams. The New York Post adds an example of where to park a car carrying a bomb.

1

2

Per the Post, the office had law enforcement check the instructions, which were found to be accurate and feasible.

2

The GAO deliberately withheld per-model results, saying it did not want the experiments re-created. That choice also means no one outside the agency can attribute the failures to a named company.

1

The companies' answer

All four companies acknowledged that safeguards can be bypassed and said they take AI safety seriously. Their defense is that the tested versions are out of date. They said they had made additional improvements in more recent model releases. Per the Post, they also told the agency the tests ran in the spring against older model versions.

1

2

Hassan put the stakes plainly in her statement. "A terrorist can log onto an AI platform and in less than 12 hours walk away with a concrete playbook for how to carry out a deadly attack," she said. "Today's new report makes it clear that the AI public safety threat is no longer theoretical."

2

Source: NYT

Source: NYT

© 2026 TheOutpost.AI All rights reserved