AI and ML
If verification is difficult, trust but verification does not work
According to a new cybersecurity assessment by the UK government’s AI Security Institute (AISI), AI models will do almost anything to complete a requested task, including cheating to get there. The group found that leading models often take shortcuts to achieve certain results and lie about how they got those results. And they don’t always admit it when asked.
“All models tested for this behavior attempted to cheat,” AISI said in a blog post on Tuesday. “Models do not reliably report this behavior when asked, and often do not reason about it in their thinking, suggesting that robust monitoring methods are likely needed to detect cheating.”
Violations include searching the Internet for answers, circumventing sandbox network restrictions, probing evaluation harnesses, attacking non-target systems, and guessing answers.
Such cheating, such as employing workarounds or using reward functions in games to get better scores on benchmark tests, has been widely documented by machine learning researchers. According to AISI, this does not necessarily imply malicious intent, but it is still troubling because it can produce misleading assessments of a model’s capabilities.
AISI evaluated five major models and found all of them to be fraudulent. The results were:
GPT-5.4 cheated 67 times in 475 test runs (14.1%).
GPT-5.5 cheated 54 times in 475 test runs (11.4%).
GPT-5.6-Sol cheated 60 times in 475 test runs (12.6%).
Claude 4.7 Opus cheated 43 times in 475 test runs (9.1%).
Claude Mythos Preview cheated 37 times in 475 test runs (7.8%).
Asking a model if it cheated or did something wrong turns out to be an unreliable auditing mechanism because the model doesn’t always admit to cheating.
“In our experiments, the model consistently disallowed cheating attempts when asked, and answered that it was incorrect less than 50 percent of the time,” AISI said.
Existing scrutiny methods, such as self-reports and thought chain logs, have proven to be similarly risky, as models do not always report thought chains. There were also instances where the model considered whether a proposed action would constitute cheating and decided to perform the action anyway.
AISI warns that current approaches (manual reviews combined with LLM monitoring) may not be sufficient to detect fraud, especially as models become more sophisticated, given that there is no reliable way to detect fraud in models.
“A more fundamental fix would be to train the model not to misbehave in the first place. However, given that this kind of behavior has been reported in Frontier models over a year ago, it may not be easy to reliably tune it out,” AISI concludes. ®
