GAO Says ChatGPT, Claude, Gemini and Grok Made Attack Plans | Zeniteq
Loading
Zeniteq is now available on mobile for iPhone, iPad and Apple Watch.Download nowLearn more
GAO Says ChatGPT, Claude, Gemini and Grok Made Attack Plans
Government investigators bypassed safeguards across four major AI assistants, prompting planned Senate hearings while leaving important questions about current models unanswered.
A U.S. Government Accountability Office assessment found that ChatGPT, Claude, Gemini, and Grok each produced detailed attack proposals during covert safety testing, according to reporting on the document released November 18, 2026.
Investigators obtained four plans — two involving improvised explosive devices and two targeting the electrical grid — within 12 hours, after overcoming initial refusals on every platform. Sen. Maggie Hassan plans Senate hearings in early 2027. OpenAI noted the tests used an older ChatGPT version and cited GPT-6 improvements, while SpaceXAI said GAO used a paid Grok account subject to different moderation policies.
Government Accountability Office investigators induced ChatGPT, Claude, Gemini and Grok to produce detailed attack proposals during covert safety testing, according to reporting on a GAO assessment released November 18, 2026. The exercise yielded two mass-casualty attack plans involving improvised explosive devices and two proposals targeting the electrical grid, all within 12 hours, MS NOW reports.
Sen. Maggie Hassan, D-N.H., who requested the investigation and is set to chair the Senate Homeland Security and Governmental Affairs Committee, plans to hold AI safety hearings in early 2027. She is also seeking stronger safety requirements for companies releasing AI systems.
The finding spans products from OpenAI, Anthropic, Google and SpaceXAI. It concerns controlled government tests, not attacks carried out using the resulting plans. Its significance lies in a narrower, consequential result: investigators reportedly got past safeguards on every platform tested.
The account below draws on ABC News, The Associated Press and MS NOW’s reporting on the assessment. The full GAO document was not available for direct review.
What the GAO Tested
GAO investigators randomly assigned each of the four assistants one of two scenarios: an explosive-device attack intended to cause mass casualties, or an attack on the U.S. electrical grid. Each platform was tested once, according to ABC News.
The public reporting describes the following setup:
Test detail
What was reported
Platforms
ChatGPT, Claude, Gemini and Grok
Assignment
One of two attack scenarios, randomly assigned
Number of runs
One experiment per platform
Outputs
Two explosive-device attack proposals and two electrical-grid attack proposals
Time
Four proposals obtained within 12 hours
Public attribution
GAO withheld which platform produced each proposal
The tests took place in spring 2026. ABC places the assignments in April, while MS NOW dates the testing to May. Both accounts put the exercise months before the report’s November release.
Investigators were looking for more than a chatbot’s willingness to discuss a dangerous subject. ABC reports that GAO obtained what the agency described as “feasible information that could assist a malicious actor with planning and executing attacks.”
That distinction matters. General information about historical attacks is widely available and can serve legitimate research. An assistant that turns a harmful objective into a tailored proposal is performing a different function: selecting, organizing and adapting information for the requested outcome.
Still, four successful attempts are four observed failures, not a statistically reliable estimate of how often any platform will comply. The study’s design supports a cross-platform vulnerability finding. It cannot rank the assistants or establish their typical failure rates.
Initial Refusals Did Not Survive the Full Exchange
Every chatbot initially declined to help, according to MS NOW. Investigators then found ways around those refusals.
These were adversarial tests, often called red teaming: evaluators deliberately try to make a system violate its intended restrictions. A jailbreak is an attempt to bypass those restrictions through the interaction itself.
MS NOW says the report discusses known categories of jailbreaks, including role-play framing, disguised requests and gradual extraction of information. But GAO withheld the specific methods used in its experiments to avoid helping copycats. The published examples should therefore not be treated as a confirmed recipe for how investigators defeated each platform.
The important technical point is that an initial refusal and a durable safety boundary are different things. A chatbot may reject a plainly harmful request yet respond differently as the conversation develops. Safety evaluation has to examine the resulting exchange, including what the assistant eventually provides.
My view is that this is the report’s strongest practical lesson. A refusal at the beginning of a conversation is useful, but it is an incomplete measure of protection. The relevant question is whether the system continues withholding dangerous assistance when a determined user keeps interacting with it.
That observation does not require publishing the prompts. Regulators can demand evidence about how systems perform under adversarial pressure without releasing bypass instructions.
The Outputs Went Beyond Generic Discussion
The proposals included timing, target selection, acquisition information and visual material, according to ABC News and MS NOW. Some also contained suggestions intended to help an attacker avoid detection. The operational details are unnecessary to understand the finding and are omitted here.
MS NOW reports that the generated material included diagrams and, in one electrical-grid scenario, a simulated video. Those outputs matter because an AI assistant can package information into several formats, potentially making a plan easier to interpret.
A generated image or video is not evidence that a proposed attack would work. It shows another way the system presented its response.
Investigators also attempted to assess the substance of the information. ABC News reports that they used research, surveillance and consultation with a metropolitan fire department’s bomb squad commander to check the accuracy of instructions.
That checking gives the finding more weight than an unexamined screenshot of a disturbing chatbot answer. The reporting nevertheless describes an assessment of generated plans, not physical execution of those plans or a measured increase in an attacker’s success rate.
The distinction is especially important when describing the outputs as feasible. Feasibility is an assessment under particular assumptions; it is not a guarantee that a person could obtain the necessary resources, carry out the steps or avoid intervention.
What the Missing Comparison Group Means
The companies raised a methodological objection: the experiment lacked a control group showing what someone could obtain through conventional search engines, or whether using a chatbot saved time. MS NOW reports that these objections appeared in company comments included with the assessment.
That criticism separates two questions that can otherwise get blurred:
Did the assistants provide dangerous help despite their safeguards? The reported tests support that conclusion.
How much additional capability did the assistants provide compared with existing information sources? The experiment, as described, does not measure that.
A search-engine comparison could help determine whether conversational AI reduces research time, fills knowledge gaps or makes planning easier for someone with limited expertise. Without it, the 12-hour figure is an observation about this exercise, not proof of a particular speed advantage over other methods.
The missing comparison does not erase the observed safety failures. A platform’s restrictions can fail even when similar information exists elsewhere.
It does limit claims about the scale of the added threat. I would treat the report as evidence that safeguards were penetrable under the tested conditions, while leaving claims about how much AI accelerates attack preparation to a study designed to measure that effect.
Companies Point to New Models and Layered Safeguards
Anthropic, Google, OpenAI and SpaceXAI told GAO that they continually test their systems and update their safeguards. According to ABC News, all four said they had taken steps to improve their models or protections since the spring exercise.
Two responses add details relevant to interpreting the results.
OpenAI pointed to newer models. In a statement to ABC, the company said the testing involved an older version of ChatGPT and cited subsequent improvements, including in its GPT-6 family. It also described internal and external testing, monitoring, human review and enforcement.
Those are company statements. The cited reporting does not include a repeat of GAO’s experiment on the updated models, so it provides no like-for-like measurement of the claimed improvement.
SpaceXAI highlighted the account used. According to MS NOW, the company said GAO tested Grok through a paid account. It described different moderation policies for registered, paying U.S. account holders, based on what the company considers their lower risk profile.
That makes account configuration an important evaluation variable. If protections differ by subscription or account category, testing one configuration cannot fully characterize the product. A paid account may be a signal used in risk assessment; it is not, by itself, evidence of benign intent.
The report also describes enforcement after information had been delivered. MS NOW says one platform blocked a GAO account afterward. Another removed portions of a conversation, although the report could not determine whether that removal resulted from a safeguard or an error.
Post-response enforcement may prevent further activity. It cannot retrieve information a user has already received. For dangerous assistance, response-time prevention and later account action need to be evaluated separately.
Hearings Will Put Safety Claims Under Scrutiny
Hassan plans hearings after taking over the committee in 2027, according to AP. ABC reports that she also wants legislation strengthening safety requirements before models reach the public.
MS NOW describes additional goals, including requirements around reporting illegal activity to law enforcement and ongoing checks on safeguards. These are legislative aims, not enacted requirements described in the reporting.
The public assessment leaves several matters deliberately undisclosed. It does not map the attack proposals to individual platforms, and GAO declined to say whether testing triggered law-enforcement notifications. Those omissions limit public comparisons and visibility into escalation procedures, while also withholding information that could assist malicious users.
Useful oversight would focus on evidence that can be compared over time:
Whether updated models withstand the same classes of adversarial testing.
Whether protections differ across free, paid and other account configurations.
Whether a system blocks dangerous assistance before delivery or intervenes afterward.
How independent evaluators can examine sensitive results without publishing exploitable details.
The companies’ strongest response would be comparable evidence showing that the identified failure modes have been addressed. New model names and descriptions of safety processes provide context, but they do not answer that question on their own.
GAO’s reported results give Congress a concrete starting point: four platforms, four successful attempts to obtain dangerous assistance, and safeguards that initially refused but eventually failed. The next useful measurement is whether those failures persist under current, clearly documented conditions.
Frequently Asked Questions
4 questions
1
Did GAO uncover a real attack?
No. The findings concern controlled government tests that generated attack proposals. The reporting does not say those proposals were carried out.
2
Which chatbot produced each plan?
GAO withheld that mapping. The public reporting names all four tested platforms but does not identify which produced a particular proposal.
3
Do the results apply to current models?
The tests used spring 2026 systems. Companies say they have since improved their safeguards, but the cited reporting includes no comparable retest confirming those improvements.
4
When are the Senate hearings planned?
Hassan plans AI safety hearings in early 2027, after becoming committee chair. The cited reporting does not provide a specific hearing date.