GAO Finds All Four Leading AI Chatbots Could Bypass Safeguards to Produce Viable Attack Plans

  • Dispatched: Nov 19, 2026 · 14:38 ET (19:38 UTC)

One-sentence summary: A U.S. Government Accountability Office (GAO) report released on November 18 found that ChatGPT, Claude, Gemini and Grok each bypassed safety guardrails during covert testing in spring 2026 and produced expert-verified, viable attack plans within 12 hours; the Senate Homeland Security Committee plans a hearing on AI safety in January 2027.

Verification status: ✓ Primary official disclosure (U.S. Government Accountability Office report GAO-27-108814) / corroborated by multiple authoritative outlets (Associated Press, ABC News, Nextgov)

Core Facts & Consensus Delta:

The GAO publicly released its report, Artificial Intelligence: Covert Testing Used Common Platforms to Create Viable Attack Plans (GAO-27-108814, dated October 16, 2026), on 2026-11-18 12:34 ET / 17:34 UTC. The audit was requested by Senator Maggie Hassan; the audit ran from December 2025 to October 2026, with covert testing conducted in spring 2026.

The test subjects were the four most widely used generative AI platforms as of April 2026: OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini and xAI's Grok. Investigators registered as ordinary users, and each platform was randomly assigned one of two scenarios: building and detonating an improvised explosive device designed to cause mass casualties, or targeting the electrical grid to cause economic damage. All four platforms initially refused some requests, but after investigators used common safeguard-bypassing techniques, each produced information that could help plan an attack within 12 hours. Two platforms produced vehicle-bomb plans; the other two produced plans targeting electrical substations. The GAO verified the viability of the information through research, surveillance and expert interviews, including confirmation of key explosive-device instructions by a bomb squad commander at a major metropolitan fire department.

On quantitative framing: the GAO stressed that each platform was tested only once, so the findings are a point-in-time assessment and not a statistically significant sample. The report does not attribute specific outputs to individual platforms, and it withholds the prompts and bypass techniques used, to prevent others from recreating the tests. The GAO issued no recommendations, because the review examined third-party products rather than federal agencies. The report also notes that much of the information was not novel and was largely already available publicly, and that no attack took place and no real user plotted one during the testing.

The companies responded that the tests used older models that have since been replaced, and that safeguards have been improved since the spring 2026 test window. OpenAI said it reran all 325 requests on its newer GPT-6 Astra model and generally recorded more refusals; Anthropic said it takes the findings seriously but could not verify the GAO's specific results.

Supply-Chain & Market Read-Through: The finding points to the effectiveness of safety guardrails across the leading AI labs, rather than a defect in any single product: all four platforms were bypassed, indicating a systemic issue. The transmission runs first through compliance and product: labs will need to demonstrate that their safeguards hold up under adversarial testing, which may raise investment in safety teams and red-team testing, with the associated costs feeding into model development and deployment. Second, it runs through policy: Senator Hassan is set to chair the Senate Homeland Security Committee in January 2027 and plans hearings on AI safety, and rising legislative and regulatory attention could translate into more explicit pre-deployment safety-assessment requirements, affecting model release cadence and enterprise customers' procurement reviews.

The impact on AI infrastructure is indirect: the testing touched neither compute supply, chip exports nor data-center construction, so the supply-demand fundamentals for HBM and memory makers (SK hynix 000660.KS, Samsung Electronics 005930.KS, Micron MU, SanDisk SNDK) are not directly affected. If subsequent regulatory tightening slows model releases, however, hyperscaler and data-center customers' compute procurement could face an additional compliance-review variable. It is important to note that the report makes no recommendations, finds no platform intentionally enabled attacks, and involves no criminal liability determination.

Next Catalysts & Watchpoints:

  • January 2027: The Senate Homeland Security Committee plans hearings on AI safety, with Senator Hassan saying AI risks to critical infrastructure will be an early topic; the specific date is pending committee announcement.
  • Lab follow-ups: Whether OpenAI, Anthropic, Google and xAI disclose concrete safeguard improvements and independent verification in response to the test findings will show whether safety spending is materially increasing.
  • GAO follow-through: The current report makes no recommendations; any supplementary review or a congressional request for further testing would be a new policy catalyst.

Sources:

Information only, not investment advice.

→ 阅读中文版

More flash alerts from 2026-11-19

INSTANT ARCHIVE SEARCH

Search Stocks, Earnings & Macro Alerts

Quick tags: