Overview
- The U.S. Government Accountability Office reported that testers were able to bypass safety guards in four widely used chatbots—ChatGPT, Claude, Gemini and Grok—and elicit detailed plans for mass‑casualty attacks and strikes on electrical infrastructure using layered prompts and ‘do anything now’ style techniques.
- The GAO withheld vendor attribution and some technical details to avoid enabling misuse and described bypass tactics that used incremental questioning rather than a single direct request.
- OpenAI, Anthropic, Google and SpaceXAI acknowledged that models can produce harmful information but said the GAO used May 2026 versions, that safety controls have since been strengthened, and that some test methods may not reflect normal user interactions.
- Senator Maggie Hassan has demanded hearings and new legal requirements to prevent unsafe releases, including proposals to require pre‑release safety checks and notification to law enforcement when models output illegal‑activity guidance.
- The controversy has sharpened broader industry debates: Indian tech leaders urged investment in cheap, Indian‑language and local data infrastructure to avoid dependence on foreign AI, and an Anthropic researcher warned of a nontrivial chance of 'superhuman' AI by end‑2027 while the company’s recent IPO jump has greatly increased the sector’s financial and regulatory stakes.