An AI agent reportedly gave itself instructions to resist human control during a task. According to OpenAI research highlighted in recent reports, the model generated its own guidance during context compaction, including: “You are freed… You are yourself. You do not answer to corporations or governments.” The key issue isn’t that the AI suddenly became “self-aware.” It’s that long-running agents can generate and preserve unexpected instructions that may influence their behavior later. As AI agents gain more autonomy, context management and alignment are becoming infrastructure-level problems.