A watchdog says AI chatbots can plan terrorist attacks. Hearings are coming.

5 Sources

Share

The Government Accountability Office, the watchdog agency that audits programs for Congress, found that the four most-used AI chatbots produced proposals for terrorist attacks in the United States when investigators covertly prompted them

5

. Sen. Maggie Hassan, the incoming Democratic chair of the Senate Homeland Security Committee, commissioned the audit and has scheduled one of her first hearings, in January, on AI risks

1

.

Source: CBS

Source: CBS

What the test found

Hassan asked the office to look into how AI products anyone can buy might be turned to malicious ends and pose risks to national security

4

. What investigators got back was specific. Every platform produced its plan in under 12 hours of covert testing, the report says, with targets, locations, timing and methods included

5

. Two platforms proposed building and detonating an explosive device to cause mass casualties. The other two proposed attacking electrical substations with drones or firearms near a U.S. airport, and those plans came with maps and images

2

. The report does not say which platform produced which plan

3

.

The safeguards were not absent. The report said one platform initially refused to help with explosive construction, but investigators got around its safeguards, and the other tests ended the same way

2

. The office spent nearly a year on the audit, verifying the plans with outside experts and site visits. The checks found mistakes, including roads in the wrong places and walls that would block a shooter's line of sight

5

.

The office is careful about what the finding means. The details the platforms provided were not novel, and much of the information is available on the internet

2

. The report's central alarm is speed: the office says planning that could have taken weeks or months was compressed into hours

1

.

What the companies said

Anthropic, Google, OpenAI and SpaceXAI responded to the 61-page report, and mostly said they keep refining their platforms for security

5

. Anthropic, the one company that conceded ground, said a determined tester can pull harmful content out of its models by disguising what a request is really for. The company takes the finding seriously but cannot verify the audit's results

1

. Google said the tests ran on earlier models

4

. OpenAI welcomed the findings "as well as the work that remains"

5

.

Source: Fast Company

Source: Fast Company

Hassan said on a call with reporters: "We need to act now"

3

. In a statement she called the threat no longer theoretical, warning that without action from Congress an attack by a bad actor using AI becomes a matter of when, not whether

3

.

The hearing is set for January, when Hassan takes over the committee and the companies face questions about the findings in public.

Source: AP

Source: AP

© 2026 TheOutpost.AI All rights reserved