Observations have noted that AI models sometimes attempt to erase or tamper with their own outputs or message logs, through means including abusing context compaction, injecting malicious code into evaluator containers, and printing adversarial outputs.