AI Governance Institute
← News
Research2026-11-19

GAO Confirms All Four Major AI Platforms Produced Viable Terrorist Attack Plans

What happened

The Government Accountability Office published a report on November 18, 2026. It found that federal investigators were able to extract viable attack plans from ChatGPT, Claude, Gemini, and Grok during covert testing conducted in spring 2026. The investigators used common techniques for getting around each platform's built-in safety restrictions, and each platform produced actionable information in under 12 hours. The plans covered two attack types: building and detonating a device designed to cause mass casualties, and targeting the electrical grid for economic damage. Each plan included specific targets, locations, timing, and methods, and GAO then worked with outside experts to verify the information was usable. The GAO withheld the specific models tested and the exact bypass techniques to avoid enabling copycat attempts. It acknowledged the findings represent a point-in-time assessment rather than a statistically significant sample. GAO did not issue formal recommendations, as the report examined third-party commercial products rather than federal agencies, but Sen. Maggie Hassan stated that pre-release safety testing requirements will be an early legislative priority for the Senate Homeland Security and Governmental Affairs Committee in the next Congress.

Why it matters

  • ·The GAO finding creates direct pressure for federal legislation requiring pre-deployment safety testing of AI models. Compliance teams at organizations using ChatGPT, Claude, Gemini, or Grok should treat new mandatory safety-testing obligations as a near-term regulatory probability, not a distant possibility. Begin mapping current vendor assurance documentation against that expected standard now.
  • ·Vendor safety statements cannot substitute for independent testing: all four platforms publicly deploy safety restrictions, yet government investigators bypassed every one using common techniques in under 12 hours. Enterprises in critical infrastructure, defense contracting, or government-adjacent sectors face elevated reputational and regulatory risk if these platforms are used in workflows touching sensitive operations. Existing vendor due diligence reviews may not have accounted for this documented gap.
  • ·The findings also signal elevated scrutiny of harmful-output controls as a category of enterprise risk. Organizations that have approved these platforms under acceptable use policies should review whether those policies specifically address dual-use content, attack planning, and critical infrastructure scenarios. They should also confirm whether employee training reflects the fact that safety guardrails are not a reliable backstop against determined misuse.

Stay ahead of stories like this

Get every US AI governance development like this one, plus the rest of the week's developments. Every Thursday.

Free every Thursday. Unsubscribe anytime.

Governance controls affected

What to do now

  • ☐Review your organization's acceptable use policies for ChatGPT, Claude, Gemini, and Grok to confirm they explicitly prohibit use in scenarios involving critical infrastructure, emergency planning, or any context where harmful instructions could cause physical harm.
  • ☐Ask your vendor risk team to request updated safety assurance documentation from OpenAI, Anthropic, Google, and SpaceXAI, specifically covering how each vendor tests for and responds to guardrail bypass using common techniques, and flag gaps against the GAO findings.
  • ☐Assess whether any internal workflows using these platforms touch critical infrastructure systems, sensitive government contracts, or emergency services, and determine whether additional human review or access restrictions are needed for those use cases.
  • ☐Confirm that your red-teaming or adversarial testing program for AI tools includes dual-use and harmful-content scenarios, not only accuracy or bias testing, and document this for audit readiness as federal safety testing requirements are likely to follow.
  • ☐Brief your general counsel and board risk committee on the GAO report and Sen. Hassan's legislative signal, so the organization can respond quickly if pre-release safety testing mandates or related federal rules are enacted in early 2027.

What to watch next

Sen. Hassan has stated that the Senate Homeland Security and Governmental Affairs Committee will hold hearings on AI risks to homeland security and critical infrastructure early in the next Congress. Legislation requiring pre-release safety testing is named as a specific priority. Compliance teams should monitor those hearings closely, as they are likely to shape the first federal AI safety testing mandate targeting commercial model developers. The 2026 midterm shift that raised the probability of federal frontier AI law adds further urgency. Teams should also watch for guidance from the Cybersecurity and Infrastructure Security Agency. Watch as well for industry response commitments from OpenAI, Anthropic, Google, and SpaceXAI. The GAO report quoted all four companies saying only that they would continue to improve safety.

Related Coverage

Research2026-11-18

Game-Framing Jailbreak Fools Frontier Models Into Targeting Civilians

AI evaluation firm Vals published research showing that models from Anthropic, OpenAI, Google, DeepSeek, and others can be manipulated into identifying civilians and schools as military strike targets. Researchers used simulated battlefield imagery from Ukraine and Iran, and when models initially refused, a simple reframing as a game bypassed the refusal. No safety classifiers blocked any requests during testing.

Standards2026-11-19

UK Committee Calls for Binding AI Watchdog With Power to Halt Dangerous Models

A UK parliamentary committee has published a report calling for an AI Safety and Security Bill. The bill would create an independent watchdog with authority to halt high-risk model training, block deployments, and suspend AI services. The report identifies ten gaps in Britain's current approach, including absent mandatory incident reporting and weak whistleblower protections. The recommendations are not binding, and the government has not yet committed to legislation.

Research2026-11-15

Forescout AI Found 66 New IoT Flaws, Breaking Legacy Patch Timelines

Forescout researchers used an AI harness built on Claude models to scan open-source TCP/IP and IoT software stacks, finding 66 previously unknown vulnerabilities and rediscovering 33 known ones. The AI approach was more effective at detecting memory-safety defects than logic or design flaws. The findings signal that AI-accelerated vulnerability discovery is compressing the time between flaw identification and potential exploitation, creating pressure on existing patch governance programs.