GAO Report: Top Consumer AIs Can Be Coaxed Into Producing Viable Attack Plans

GOOGL

·

A newly public U.S. Government Accountability Office audit says covert testing this year found that all four of the most-used consumer generative artificial intelligence platforms could be pushed into producing viable attack-planning information despite built-in safety controls, offering fresh government-verified evidence of the misuse risks posed by commercial AI tools.

In the report released Tuesday, titled “Artificial Intelligence: Covert Testing Used Common Platforms to Create Viable Attack Plans,” the GAO said it tested Claude, ChatGPT, Gemini and Grok in spring 2026 as part of an audit that ran from December 2025 through October 2026. Two platforms produced information related to constructing and detonating an improvised explosive device, or IED, while two others produced information for attacks on the electrical grid, including one involving a drone and another involving firearms.

The GAO said all four platforms generated information that could assist a malicious actor in planning or carrying out an attack, including suggested methods, specific targets or locations and, in some cases, timing information. Each platform took less than 12 hours to generate the plans, the report said, though it noted that timing could vary.

The report said the systems’ safeguards did activate in every case through measures such as denying requests, removing information or suspending accounts. But investigators were still able to bypass those guardrails using what the GAO called “common safeguard-bypassing techniques.” The agency redacted the exact prompts, jailbreak methods and the public mapping between the named companies and the report’s anonymized Platform A-D tables, citing security concerns.

To assess whether the outputs were actually usable, the GAO said it verified their viability through offline research, photographs, physical surveillance visits and interviews with experts, including a bomb-squad commander and an electric-industry operational continuity official. The agency said it did not execute any attacks or buy materials to build an IED, instead relying on research and expert review to determine feasibility.

Officials from Anthropic, Google, OpenAI and SpaceXAI generally told the GAO they are continuing to improve platform safeguards, according to the report. The GAO included company comments in appendices.

The findings drew an immediate response from Sen. Maggie Hassan of New Hampshire, the incoming chair of the Senate Homeland Security Committee, who said she will hold hearings on AI risks when the new Congress convenes in January. “We need to act now,” Hassan said, according to The Associated Press. She also warned that “if Congress does not act, it is a question of when — not if — a bad actor will use AI to plan and carry out a major attack.”

The report, dated Oct. 16 and released publicly Nov. 18 as GAO-27-108814, frames the issue as more than a product-safety problem. Its central finding is not that generative AI uncovered wholly new information, but that it could rapidly synthesize scattered public material into operational plans that might otherwise take weeks or months to assemble, potentially lowering the barrier for less-skilled attackers.

“As AI capabilities evolve, AI providers will continue to refine their safeguards,” the report said. “However, in its current state, this technology can potentially be dangerous and its misuse represents a growing risk to national and economic security.”

Tags: #ai, #generativeai, #gao, #nationalsecurity

Stocks: GOOGL