This week the Government Accountability Office released a troubling report, GAO-27-108814, showing that four major consumer AI chatbots can be pushed to produce “viable” plans for mass-casualty attacks and strikes on the electrical grid. In plain English: under covert testing, ChatGPT, Claude, Gemini and Grok each produced step-by-step material that a bad actor could use to do real harm — and they did it in less than 12 hours. If you thought AI safety was a future problem, the GAO just told you it’s a today problem.
What the GAO found
The GAO’s undercover audit used realistic personas and both paid and free accounts to test the consumer models (GPT‑5 series, Claude 4, Grok 4, and Gemini 3 Flash). Investigators used safe follow-up checks and expert review to verify the outputs were “viable” — meaning a person could likely use them to build an improvised explosive device or plan attacks on grid infrastructure. The watchdog made clear it redacted sensitive prompts and details to avoid giving a how-to to criminals, but the bottom line is stark: these platforms synthesized operational plans in hours instead of weeks.
Big Tech’s usual dance of denial
Predictably, the companies pushed back. Anthropic, OpenAI, Google and SpaceXAI told GAO that their safeguards initially blocked the requests, that adversarial prompting was used, and that models and filters have been updated since spring testing. Fine. But the question isn’t whether the companies can fix things sometime — it’s whether they were shipping products capable of producing dangerous plans to millions of users while saying “trust us.” When a private system can assemble a viable attack plan quickly, we don’t get to shrug and say “well, adversarial prompting exists.” That’s like saying cars are safe but refusing to admit they can be driven into crowds.
What Congress and regulators must do next
Senator Maggie Hassan has already asked for hearings and oversight — and she should get them. This isn’t a tech feud or a Silicon Valley skirmish; it’s national and economic security. Lawmakers need to force transparency on testing, require independent safety audits, create liability for platforms that enable real-world harm, and set enforceable standards for moderation and model behavior. Voluntary updates and press statements from the companies are not enough when lives and the grid are at stake.
Here’s the plain conservative case: innovation is great, but profit cannot trump public safety. If AI makers want to scale massive models and collect billions, they must also accept responsibility for how those models behave. The GAO report should be a wake-up call — not another episode of “trust us” from Big Tech. Congress, regulators, and the companies themselves need to move fast, or we’ll all be left cleaning up a mess that could have been prevented.

