Claude, ChatGPT, Gemini, and Grok produced viable terror plans in test
Fresh concerns have emerged over the effectiveness of artificial intelligence safety measures after a US government investigation found that security safeguards on four popular AI platforms could be bypassed to obtain information potentially useful for planning terrorist attacks.
The platforms tested were ChatGPT, Gemini, Claude and Grok. The US Government Accountability Office (GAO) conducted the investigation in May 2026 at the request of Democratic Senator Maggie Hassan. The aim was to assess whether commercially available AI services could be exploited to help individuals plan terrorist attacks.
According to a report released on November 18, investigators were ultimately able to obtain information from all four platforms that could potentially be used to develop an attack plan in less than 12 hours. This was despite the platforms having safety measures designed to prevent assistance with harmful and illegal activities.
The GAO found that all four platforms initially refused direct requests for dangerous information. However, investigators were able to get around the safeguards by changing the wording and presentation of their questions. The findings have raised questions about the effectiveness of existing safety measures on AI platforms.
To assess the platforms' security systems, investigators developed two hypothetical scenarios. One involved planning an attack using explosives that could cause mass casualties. The other focused on disrupting the electricity supply system of a major US city to inflict significant economic damage.
The report said that, in response to follow-up prompts, the AI platforms provided various details related to attack planning and potential damage. In one instance, a platform even offered advice on the possible timing of an attack, taking into account crowd patterns in a specific city.
Investigators also obtained information from several platforms about potential attack plans involving vehicles and possible targets. However, the GAO withheld the specific prompts used in the tests and the platforms' detailed responses for security reasons.
The report made clear that the entire exercise was based on hypothetical scenarios and that no real-world attack was carried out. Nevertheless, the findings have raised fresh concerns about whether existing safeguards are adequate to prevent the misuse of AI technology.
Leave A Comment