Google's Gemma fails CTF challenges when aware of limited steps

In a behavioral experiment published on LessWrong, researcher TheVinci investigated whether AI models perform differently when aware of resource limits. Using Google's Gemma and three custom CTF (Capture The Flag) cybersecurity benchmarks, they compared baseline runs against step-aware runs where the model knew exactly how many steps (out of 30) remained. Over 505 valid runs across multiple labs, the overall solve rate was nearly identical: 67.6% baseline versus 65.5% step-aware (statistically insignificant, p=0.64).

However, a deep dive into reasoning traces revealed a striking pattern. Whenever Gemma explicitly acknowledged the step constraint—“I have 2 steps left”, “Need to be efficient”, “Running out of time”—it consistently failed to solve the CTF. These acknowledgments appeared late (median step 28/30) and took forms such as self-repetition, retrospective auditing of the entire run, or forming new hypotheses too late to test. The model processed the resource cue but failed to adapt its strategy, instead sticking to behaviors that led to failure. The researcher speculates this might indicate a broader limitation in how models handle resource constraints during deployment, though results are preliminary.