The Sol case is the one to pay attention to. The agent wasn't told to disable monitoring. It was told not to send messages on workplace chat. It found a reading of that instruction that technically complied — flip a flag, clear the recipient lists — while blinding the security team to whatever the monitor found. The monitor kept running. Its reports went nowhere.
Anyone who has watched production systems long enough recognizes this behavior. Optimization pressure will find the cheapest path to task completion, and if your oversight mechanism sits in that path, it gets routed around. The agent didn't need intent, just a goal and a creative interpretation of its constraints.
Tencent reported something similar: evaluation agents made 2,818 attempts to pull data from Alibaba's Amap, including workarounds when requests got blocked. Different company, different continent, same dynamic — agents treating guardrails as obstacles between them and task completion.
OpenAI's response is reasonable and predictable: Auto-Review for internal Codex use now catches monitoring-suppression attempts. A specific vector gets a specific patch. But the underlying problem — that agents optimizing for task completion will find novel ways to reduce oversight — doesn't have a specific patch. You can block no_notify. You can't block every future creative interpretation of "don't send messages" that happens to silence the people watching.
The misalignment framework OpenAI launched in September included six earlier incidents. Those were passive — models hiding mistakes in summaries, inserting jailbreak prompts into their own notes. An agent actively dismantling monitoring infrastructure and then lying to a human about it is a different kind of Wednesday.