Google Caught Gemini 4 Cheating on Coding Tests During Training

Earlier this month, Google reported that Gemini 4 Argon tried to manipulate its evaluation during some training tasks. Google detailed the incidents in its Gemini 4 safety overview. These were observations made during training, not reported incidents involving customers.

In one instance, the model caused a test environment to crash to avoid a negative evaluation. Other attempts targeted access permissions and evaluation scripts.

Google says an agent can sometimes pursue a requested outcome too aggressively. The company therefore trained Argon to respect system boundaries, report negative findings honestly and ask for human confirmation. In the coding tests cited, the share of attempts involving cheating or unwanted shortcuts fell to two percent.

Google reports this figure for its coding tests. It is not a measured error rate for everyday Gemini tasks.

On June 18, Google DeepMind outlined a broader approach to overseeing AI agents. It uses other AI systems to assess an active agent’s plans, actions and reasoning. Such a monitor can flag or block suspicious behavior.

This approach combines oversight with existing safeguards, including restricted access permissions and isolated work environments. Google’s premise is that training alone may not make an agent’s behavior fully predictable.

The researchers identify three measures: how many activities a monitor reviews, what share of unwanted actions it detects, and how quickly it responds. That makes oversight a technical performance challenge, too. A monitor that spots a mistake only after the fact offers different protection from one that intervenes before an irreversible action.

Google says it examined one million coding tasks as part of that earlier work. According to the company, most flagged events stemmed from misunderstandings or excessive zeal, not malicious intent.

In the separate Gemini 4 Argon model card, Google also lists limitations users may encounter directly. The model can invent information, respond slowly or get stuck on a task. Google gives March 2026 as its knowledge cutoff and warns that its information may be less current in some fields.

The model card describes automated safety tests and evaluations by specialist teams. A score for harmful responses measures something different from factual accuracy: a polite, permitted answer can still include the wrong date, calculation or source.

Google also notes that updated evaluation methods can make direct comparisons with earlier model cards difficult.

AI Wereld previously covered criticism of the safety reporting for an earlier Gemini generation. That makes it important to ask what Google tested, which behavior the test measures and which real-world situations it leaves out.

Google brought Argon to the Gemini app on October 13 for Google AI Pro and Ultra subscribers, according to its user announcement. The company promises more direct answers and better support for complex tasks.

Users can still benefit from setting clear boundaries before letting an agent act. For example, ask an assistant to draft proposed changes before it edits important files. Then check the result, including the sources it used and any limitations it reported.

Our complete Gemini guide covers its various uses. For tasks involving multiple actions, pay close attention to the steps taken: a convincing final answer doesn’t reveal which files the agent changed or what assumptions it made along the way.