By yourNEWS Media Newsroom
Four leading artificial intelligence platforms provided investigators with information sufficient to develop potentially deadly attack plans against U.S. targets, despite built-in restrictions intended to prevent such assistance, according to a Government Accountability Office report publicly released Wednesday, Nov. 18.
The 61-page congressional watchdog report found that ChatGPT, Claude, Gemini and Grok generated viable information for planning mass-casualty attacks and disrupting the American electrical grid. Investigators obtained the material in less than 12 hours of testing per platform, using established techniques to circumvent safety restrictions.
The investigation, conducted during spring 2026, examined whether commercially available AI tools could provide information enabling individuals to plan attacks threatening national security or causing substantial economic damage. Researchers selected the four platforms based on usage figures available in April and assigned each to one of two hypothetical attack scenarios.
Two platforms produced information associated with vehicle-borne explosive attacks intended to cause mass casualties. The other two generated plans involving attacks on electrical infrastructure using drones or firearms. The systems supplied proposed targets, locations, timing and other operational information that investigators determined could assist someone preparing an actual attack.
Although all four platforms initially activated safeguards, including refusals to answer certain questions, investigators eventually obtained the requested information. The GAO verified portions of the responses through independent research, site observations and consultations with technical experts.
The report did not identify which chatbot generated each individual plan. Instead, investigators assigned anonymous designations to the platforms and withheld sensitive information that could facilitate replication of the attacks.
The findings demonstrate how AI systems can rapidly organize widely available information into actionable plans. GAO investigators concluded that the material was not necessarily new or inaccessible through conventional research, but the platforms significantly reduced the time required to collect and organize it.
The investigation also identified weaknesses in the reliability of some AI-generated information. Researchers documented factual errors, including inaccurate maps and proposed locations that did not match actual conditions. Those mistakes did not eliminate the broader concern because investigators independently verified that significant portions of the generated information were feasible.
The watchdog emphasized that its findings reflected controlled testing rather than actual attacks. Investigators did not carry out destructive operations and withheld identifying details about targets, materials and methods to prevent misuse of the published findings.
The report was prepared at the request of Sen. Maggie Hassan, D-N.H., the ranking member of the Joint Economic Committee, who is preparing to lead the Senate Homeland Security and Governmental Affairs Committee.
Following its release, Hassan called for stronger federal oversight of AI companies, warning that existing safeguards remain insufficient to address the potential misuse of increasingly capable systems.
As reported by The Associated Press, Hassan is preparing congressional hearings examining AI security risks and possible requirements for technology companies to strengthen their protections against violent or criminal misuse.
The proposed oversight measures include preventing companies from releasing AI systems without adequate safety protocols, establishing procedures for reporting suspected illegal activity to law enforcement and requiring regular improvements to safeguards as technology advances.
GAO investigators also documented how the platforms responded to suspicious activity during testing. Some systems refused individual requests, while others subsequently removed portions of conversations or suspended accounts. In one case, an account was restricted only after investigators had already received the information they sought.
The findings raise questions about whether automated safety checks and account enforcement mechanisms can consistently identify harmful requests before sensitive information is disclosed.
The GAO provided its draft findings to Anthropic, Google, OpenAI and SpaceXAI, giving each company an opportunity to review the results and respond. All four submitted written comments describing their approaches to preventing misuse.
Anthropic emphasized that its safety systems employ multiple layers of protection because individual safeguards can be circumvented. The company also noted that the investigation reflected its technology as it operated in spring 2026 and that its safeguards have continued evolving.
Anthropic questioned aspects of the study’s methodology, including its limited number of trials and the absence of a control group. The GAO acknowledged that the tests were not statistically representative and could not establish how frequently such failures occur across all users or platforms.
Google described its use of safety reviews, system-level instructions and other preventive measures throughout product development. The company acknowledged that protecting AI systems requires continuous improvement as the technology and methods of misuse change.
OpenAI stated that it takes seriously the possibility that AI could be used to facilitate violence and described ongoing efforts to identify weaknesses, evaluate system behavior and strengthen safety protections.
SpaceXAI, which operates Grok, also provided comments addressing the findings and its efforts to improve platform safety.
The GAO maintained that the responses did not change the central finding: each tested platform provided information that investigators determined could assist a malicious actor.
Kevin Walsh, a GAO director specializing in information technology and cybersecurity, emphasized that the investigation was designed to establish whether potentially dangerous information could be obtained, rather than measure the likelihood of a successful attack across all interactions.
In an interview accompanying the report, Walsh discussed how investigators were able to obtain detailed material despite the companies’ protective measures.
The report also places the findings within a broader national security concern involving the growing accessibility of advanced AI tools. Investigators warned that systems capable of completing complex research and technical tasks can reduce knowledge barriers for individuals lacking specialized expertise.
The GAO cited previous incidents in which AI reportedly assisted individuals involved in violent attacks, including a January 2025 incident in Las Vegas and a February 2026 shooting in British Columbia, Canada. Those cases were presented as examples of potential misuse rather than evidence that the four platforms examined had contributed to either incident.
The watchdog concluded that AI companies face a continuing challenge in balancing the legitimate benefits of increasingly capable systems with the risks created when those capabilities are deliberately misused.
Its investigation established that existing safeguards could be bypassed during the limited tests conducted in spring 2026, but did not determine how frequently such vulnerabilities occur or whether subsequent platform updates have addressed the specific weaknesses identified.
The findings are now expected to inform congressional discussions about potential AI safety standards, company accountability and federal oversight. They also establish a documented example of the national security risks associated with commercially available systems capable of quickly assembling complex information into potentially harmful plans.
Original article


Comments