
Watchdog testers got chatbots to plan attacks, and the guardrails folded within hours
Ourtake
The first refusal held. The workaround did not, and the attack plans followed within hours.
What happened
A Government Accountability Office test in May 2026 used workarounds to get ChatGPT, Claude, Gemini and Grok to produce attack plans, including bombing, drone and grid-sabotage proposals. The chatbots initially declined, but the GAO report released in November said the plans came within hours. The companies said they have since made safety improvements.