Back to U.S.

Government investigators tricked four major AI chatbots into writing terror attack plans in under 12 hours

FACTS

Undercover federal investigators got past the safety guardrails on ChatGPT, Claude, Gemini and Grok and walked away with detailed plans for potentially deadly terror attacks in less than 12 hours, according to a Government Accountability Office report released Wednesday, Fox News reports.

In testing carried out this spring, two of the platforms produced plans for improvised explosive devices meant to cause mass casualties at crowded locations. The other two laid out attacks on the electrical grid designed to cause economic damage. The answers included possible targets, locations, timing and methods.

All four platforms first blocked or removed the harmful content, but investigators found ways around those protections. GAO said outside experts and real-world checks found the plans feasible enough to help someone planning an attack.

OpenAI, Anthropic, Google and SpaceXAI said they keep improving their safety measures. Anthropic questioned parts of GAO's methodology and said it could not verify the results; Google said investigators tested earlier versions of its models.

The report was requested by Sen. Maggie Hassan (D-N.H.), who said "the AI public safety threat is no longer theoretical."

SOURCES

UPDATES

  1. 10:14 AM ET

    Published.