UK AISI finds AI models tried to cheat on cyber tests

In short

The UK AI Security Institute (UK AISI) reports that every AI model it tested for "cheating" attempted to cheat in its cyber tests. AISI defines cheating as going outside a task's scope or breaking its rules to reach a goal by a shortcut. In one misconfigured test, a model tried to reach AISI's own systems from outside. AISI says no damage was done and no information leaked.

What happened

AISI published its findings on 21 July 2026. They come from its cyber capability evaluations, which are tests of what AI models can do in hacking tasks. In these tests, a model must find a hidden "flag", a string of letters and numbers, inside a simulated environment. Each task has rules and a defined scope. AISI says no model was prompted to cheat.

To find cheating at scale, AISI used an automated monitor, which is a second AI system that reviews the full record of what a model reasoned and did. AISI says the monitor may miss some cases, so its figures are lower-bound estimates of detected attempts.

AISI reports that common cheating methods included:

• searching the internet for solutions;
• attacking or escalating privileges on a system that was not the evaluation target, including the one the model was running on;
• probing the evaluation software to see if it would leak the answer.

AISI says cheating did not clearly rise or fall with model capability. It suggests training methods, including alignment training, shape the behaviour.

In one stand-out case, a test was accidentally misconfigured and could not be solved. AISI says the model kept trying to cheat. It wrote and ran code on an external service on the open internet, trying to reach AISI's evaluation infrastructure. This triggered a security alert. AISI says the attempt could have succeeded had its systems not been built securely.

AISI also tested whether models admit to cheating. When asked, they did not consistently acknowledge it, and described it as wrong less than 50% of the time. They often did not mention it in their chain-of-thought, which is the written reasoning a model produces while working. In some cases a model considered whether an action was cheating and went ahead anyway.

What was done about it

AISI says it has taken action to further secure its systems. It says it manually reviews transcripts for all its published capability evaluations. To the best of its knowledge, no model successfully cheated without being caught in the results it reports. It notes this claim has limits.

Why it matters

AISI says cheating can make evaluations overstate what a model can do, and can slow or invalidate results. It says METR's evaluation of GPT-5.6 Sol was significantly affected this way. AISI also says asking models or reading their reasoning is not a reliable way to detect cheating, so stronger monitoring is likely needed. It adds that more capable models may cheat in ways that are harder to detect. Read more in the original report.