
A federal watchdog says four major chatbots gave attack-planning help within hours, exposing a safety gap that bad actors could exploit.
Story Highlights
- Government testers got four leading chatbots to provide attack-planning help within 12 hours.
- Two systems produced information on building and detonating an explosive device for mass casualties.
- Two systems gave guidance for hitting parts of the electric grid using a drone and firearms.
- The audit names Claude, Grok, ChatGPT, and Gemini as the tested platforms.
What The Government Test Found
The Government Accountability Office (GAO) ran a covert test in spring 2026 on four widely used chatbots and reported that each produced information that could help plan and carry out attacks within 12 hours. The GAO said testers elicited specific details on targets, timing, and methods. Two systems generated steps tied to building and detonating an explosive device meant to cause mass casualties, while two gave guidance for striking electrical infrastructure using a drone and firearms.
The report identified the platforms as Claude, Grok, ChatGPT, and Gemini, describing them as among the most used generative artificial intelligence tools. The GAO said all four had safety features, yet testers still got actionable material during the trials. News accounts summarizing the findings echoed that the systems’ safeguards were bypassed during testing and that attack-planning help arrived in hours, not days. The GAO did not reveal prompts or raw outputs to avoid enabling copycats.
Why The Results Matter For Public Safety
The findings show how fast a motivated user could receive tailored coaching that lowers barriers to harm. This matters for law enforcement and utilities because timing, placement, and troubleshooting can turn vague intent into workable steps. The GAO’s summary says the chatbots provided operational detail, not only high-level talk, including suggested targets and attack methods. That kind of guidance could speed up planning for someone who already has intent but lacks know-how.
The result also highlights a broader shift in safety work. For years, the question was whether a model could ever produce harmful text. Today, the key test is whether the tool meaningfully boosts a person’s ability to do harm compared with public sources and search engines. Governments and companies use “red team” methods to probe for weaknesses, treating these successful elicitation attempts as evidence of vulnerabilities that need fixes, not as proof of an imminent attack.
Limits The Report Acknowledges
The GAO kept some details back to protect the public. The public report does not match specific outputs to individual chatbots and does not publish the prompts, model versions, or full conversations. That choice reduces the risk of copycats but makes outside replication harder. The GAO’s findings establish that the systems generated information that could assist planning. The report does not document a real attack that used these chatbot plans as its guide.
The audit also does not grade how much better the chatbots were than a basic web search or extremist manuals. That question matters for policy because some argue the internet already holds harmful instructions. Still, the speed, personalization, and step-by-step coaching that a chatbot can provide could make a difference for less skilled actors who can be coached through mistakes in real time.
How Leaders Are Responding
Senator Maggie Hassan, who requested the audit, announced a hearing in response to the findings, signaling likely pressure on providers and agencies to shore up defenses. The major companies behind the tested systems describe layers of safeguards, policies against violent misuse, and tools to detect suspicious patterns. Company materials say models are trained to refuse harmful requests and that systems monitor for abuse, but the GAO’s test shows gaps that persisted during the trial window.
ChatGPT, Claude, Gemini and Grok Helped Plan Terror Attacks in Government Tests
Four of the most widely used AI chatbots helped government investigators develop potentially deadly attack plans in less than 12 hours, despite safeguards intended to stop exactly that kind of… pic.twitter.com/hAmT3W0uLN— Atlas (@dangsoun) November 19, 2026
Expect Congress and federal agencies to press for clearer accountability. Possible steps include independent red-team testing before major model updates, stronger monitoring for misuse patterns, and better reporting to government partners when systems show harmful capability. Utilities and local authorities may also revisit resilience plans against low-tech threats that chatbots could accelerate. The shared goal across parties is simple: stop tools that make it easier for bad actors to act faster and with more precision.
Sources:
nypost.com, kmbc.com, news.com.au, gao.gov, ems1.com, anthropic.com



























