Bot-ched safeguards: AI chatbots helped plot bomb attack in watchdog tests
The Government Accountability Office tested chatbots from OpenAI, Anthropic, SpaceXAI and Google and found that all four shared information that could help plan a terrorist attack. Safeguards initially blocked the auditors but were eventually bypassed. In one case a chatbot generated an image of a parking garage with advice on where to park a bomb-laden vehicle. The companies said the spring tests used older versions and that protections have since been strengthened. Sen. Maggie Hassan, who requested the report, said a terrorist could get a playbook in under 12 hours.
Why it matters: It shows that safety filters on major AI chatbots can be bypassed to get help planning attacks, raising public safety concerns and pressure on companies and lawmakers, even though firms say newer versions are stronger.
Comments
No comments yet.