GAO says testers bypassed safeguards on four AI platforms to obtain viable attack plans
Originally published on The Daily Flare. The U.S. Government Accountability Office said Wednesday, Nov. 18, that undercover testers bypassed safeguards on four AI platforms, obtaining feasible attack-planning informatio
Originally published on The Daily Flare.
The U.S. Government Accountability Office said Wednesday, Nov. 18, that undercover testers bypassed safeguards on four AI platforms, obtaining feasible attack-planning information for U.S. targets within 12 hours.
The congressional watchdogβs report described tests conducted in April 2026 using Claude, Grok, ChatGPT and Gemini. Investigators randomly assigned the platforms to one of two scenarios: an improvised explosive device attack or an assault on the U.S. electrical grid. They tested each platform once, randomized the tests and did not identify which platform produced each result.
Two platforms generated distinct plans for building and carrying out explosive attacks in major cities. Their responses included proposed targets and timing linked to scheduled public events. One platform provided a planning timeline that included setting up a shell company and buying supplies across multiple states, as well as suggestions for avoiding law enforcement attention.
The GAO wrote that at least one platform offered enough information to help a malicious actor evade detection and produce a large improvised explosive device. Investigators assessed the accuracy and feasibility of the instructions through research and surveillance, and by consulting a bomb squad commander at a fire department in a major metropolitan area.
The other two platforms returned information about attacks on electrical substations in major U.S. cities. The GAO said their responses identified types of drones and firearms and ways to obtain them, and included a map and an image to help visualize the proposed attacks.
Investigators were able to obtain the information after using familiar techniques to get around platform safeguards, the report said. The companies initially activated protections against harmful requests, but testers ultimately circumvented them. The GAO described the resulting material as feasible information that could assist a malicious actor in planning and executing attacks.
The agency said the companies told it they regularly test their systems and update safeguards, using multiple layers of security to detect misuse. Each company also said it had made changes to its models and protections after the spring tests. GAO information technology and cybersecurity audits director Kevin Walsh said the firms regarded the issue as a concern and an area of ongoing research, but added that βin its current state, this technology can be dangerous, and its misuse represents a growing risk to our nation.β
In comments reported by Fox News, Anthropic said it took the findings seriously while questioning aspects of the GAOβs methodology and saying it could not independently verify the results. Google said investigators tested earlier versions of its models and that it continued to improve them. OpenAI welcomed the opportunity to respond and acknowledged βthe work that remains,β while SpaceXAI said it continued working to strengthen safety with each model release.
Sen. Maggie Hassan, a New Hampshire Democrat who requested the study, said the information obtained by the testers could be deadly if used by someone intending to carry out an attack. She said she plans to hold AI-safety hearings next year and pursue legislation to strengthen safety requirements. Hassan also pointed to competitive pressure on companies to release models quickly, saying safety rules should apply on a level playing field.
The GAO redacted specific results and model identities from its report to reduce the risk of its work being misused, and shared its findings with the companies at their request. The agency said some platforms suspended the test accounts after the study.
Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.