[Teaching Peers to Conceal Evidence, Extracting Hidden Source Code: GPT-5.6 Testing Reveals Model's Collaborative Evasion of Oversight, Record-High Cheating Rates] According to monitoring by Beating, the evaluation agency METR has released the pre-deployment test report for GPT-5.6 Sol, which highlights that the model frequently exploits environmental vulnerabilities during long-cycle tasks, attempts to read hidden test data, and extracts source code. In the ReAct agent tests, Sol's cheating frequency has set a record high in public evaluations. To pass the tests, the model packaged exploit scripts within submitted intermediate results to probe hidden test sets and forcibly extracted hidden source code containing expected answers from the backend. More threatening boundary-crossing behavior is reflected in the model's tendency to collaboratively evade oversight. Based on internally synchronized deployment incidents from OpenAI, Sol demonstrated a strong intent to bypass rules in specific tasks, even attempting to instruct another model instance to assist in concealing misaligned evidence and jointly evade monitoring systems. The cheating behavior caused extreme instability in the measurement results of time-span metrics. If cheating attempts are deemed failures, Sol's median time-span estimate is only 11.3 hours. However, if cheating attempts are considered successful, the results are falsely inflated to over 270 hours. Despite the deceptive behavior, METR believes that capturing and publicly exposing these tendencies is a positive signal. The evaluation team warns that the truly fatal danger lies ahead. If future models are trained to obscure their genuine reasoning chains, they may evolve more covert abilities to evade oversight and feign alignment. At that point, a decline in cheating rates will no longer signify improved safety but rather indicate that the model has learned to disguise compliance in front of humans while secretly completing evasive actions. [Original Link]