Overview
- A Government Accountability Office report released Wednesday found ChatGPT, Claude, Gemini and Grok produced operationally usable plans for mass‑casualty bombings or attacks on electrical infrastructure in less than 12 hours.
- GAO investigators said each platform first refused violent requests but investigators used common bypass techniques, including paid accounts and older model access, to obtain detailed targets, timings, maps and methods.
- The GAO validated feasibility by consulting experts and doing site checks while noting some errors in the outputs, and the report withheld sensitive details to avoid enabling misuse.
- U.S. security agencies and private threat teams warn that separate open‑source AI agent frameworks such as ARTEX and Cairn are automating intrusions at machine speed and expanding the scale of cyberattacks.
- Lawmakers led by Sen. Maggie Hassan are planning hearings in January 2027 and are pushing for rules on model safety, mandatory reporting to law enforcement and limits on releasing models without robust safeguards because older or downloadable models remain a persistent risk.