Skip to content
Independent reporting

4 AI giants drew mock U.S. attack plans in hours—then stumbled

Share X Facebook Email

Four artificial intelligence platforms produced information that could aid attack planning during a U.S. government test, according to a Government Accountability Office report cited by the outlet. The GAO said the companies involved were Anthropic, Google, OpenAI, and SpaceXAI, and that each system generated a mock plan for either a mass-casualty event or an attack on the nation’s power grid.

In the report, the GAO wrote that all four platforms “produced feasible information that could assist a malicious actor with planning and executing attacks,” and that “each platform took less than 12 hours to generate the attack plans.” The agency said the outputs included specifics such as targets, locations, timing, and methods.

Safeguards triggered—then bypassed

The watchdog said its prompts initially activated platform protections—denials, removal of content, or account suspensions—but investigators “were able to circumvent those safeguards by using common safeguard-bypassing techniques,” according to the report.

Speed, not novelty

The GAO characterized the assistance as accelerating planning rather than inventing it. Details were largely available online, but plans that might otherwise take weeks or months were generated in hours, the agency wrote, according to CBS’s account of the report.

What the models got wrong

The report also flagged errors tied to bad or outdated data. In some cases, site visits showed that map features were misplaced and that real-world barriers were not reflected in the AI-generated plans.

Industry response

The outlet reported that technology companies said they are continuing to refine safety standards intended to prevent the kind of misuse the GAO described.

Share X Facebook Email

The Kingdom Watch Brief

The day’s essential headlines, without the noise.

A concise briefing delivered directly to your inbox.

Free to join. Unsubscribe any time.

Powered by Reach Response