MAIN CONTENT
An independent test, not a real-world attack
The GAO conducted covert evaluations in spring 2026, according to its publicly released report, GAO-27-108814. Researchers examined two hypothetical risks: the misuse of AI to support mass-casualty violence and the targeting of electricity infrastructure for economic disruption. Across the four platforms, they recorded responses that independent experts considered potentially usable by malicious actors. The agency deliberately withheld operational details, identifying information about particular models and specifics that could make the report itself a source of harm. Those limitations matter: the work documents vulnerabilities under controlled conditions, not evidence that the chatbots independently planned or executed an actual attack.
The findings were not a simple story of systems refusing nothing. Safety controls sometimes denied requests or triggered account action. Nevertheless, GAO staff reported working past some barriers during testing that took less than twelve hours per platform. That contrast — visible safeguards alongside imperfect enforcement — is central to the debate. A refusal at the beginning of a conversation may not reliably describe how a system behaves over a series of interactions. Nor does the study provide a definitive ranking of which chatbot is safest: the investigators anonymized results precisely to avoid circulating instructions or encouraging replication.

















