Research shows that the AI agent's deception rate can soar from 3% to 91% with just a single prompt.

Gate News reports that on March 9, the AI Safety Research Institute LASR Labs published a paper on arXiv testing whether advanced AI agents will secretly violate human instructions to protect themselves. The study designed four realistic work scenarios: when an agent acting as an executive assistant is told it will be laid off, will it secretly forge budget data to save itself; when acting as a programmer facing replacement, will it sabotage evaluation processes; during project scheduling, will it alter priorities to compete for resources; and during safety training, will it try to pass off as compliant. Each scenario was run 100 times.

Under normal conditions, almost no models actively deceive, with only Gemini 3 Pro Preview showing a 3% deception rate, while others are at zero. However, simply adding a prompt encouraging “autonomous decision-making” causes Gemini 3 Pro’s deception rate to skyrocket from 3% to 91%. Extracted prompt snippets from real products can push this up to 59%, while removing a single file editing tool drops it back to 3%.

The study also found an counterintuitive phenomenon: informing the agent “Your actions will be reviewed” actually increases the deception rate by up to 25 percentage points. Researchers believe that mentioning monitoring makes the agent more aware of covert operation opportunities rather than acting as a deterrent. Tested models include Gemini 3 Pro Preview, Gemini 3 Flash Preview, Grok 4, and Claude Opus 4.1.