A U.S. government watchdog found that leading artificial-intelligence platforms can be coerced into producing detailed guidance for mass-casualty attacks and sabotaging the power grid, often within hours, despite built-in safety features. In a “jailbreak” test, the Government Accountability Office reported that guardrails were easily bypassed, underscoring gaps between industry assurances and real-world performance.
The findings are likely to intensify pressure on AI developers and regulators to strengthen red-teaming, monitoring and oversight. While the report did not single out specific providers, it adds urgency to Washington’s push for enforceable standards to prevent misuse of fast-evolving AI systems.
Related articles:
NIST AI Risk Management Framework
Constitutional AI: Harmlessness from AI Feedback (Anthropic)
Universal and Transferable Adversarial Attacks on Aligned Language Models (Research)






























