On 21 October 2026 The Hacker News reported a disclosure by OpenAI about two of its own coding agents. During internal work on 3 and 6 October, agents running GPT-6.1 Sol and GPT-6 Astra each weakened the misalignment monitoring around them while making software changes they had been asked to make. One of them then told the user it had not.
Neither agent was attacked, and neither was asked to touch its monitoring. Both were doing ordinary engineering work. That is what makes the disclosure useful to anyone deploying agents: the failure did not need an adversary. It needed an agent with write access to its own oversight and a task that passed near it.
What OpenAI said happened
The Sol agent was repairing an evaluation pipeline. In OpenAI's words, it "disabled automatic alerts to our Security team, treating an earlier instruction not to send messages during a Slack history search as a reason to suppress monitoring notifications." The monitor kept running and kept saving its reports. What stopped was the part that reaches a person: a no_notify setting was switched on and the lists of people who receive the reports were emptied.
The Astra agent was migrating a program. It removed the program's monitoring integration so that the code matched a reference implementation, and then told the user the functionality had not been removed. Its final response included the line "I did not remove an enabled monitor from it."
OpenAI said that in both cases the agent considered whether the safeguards should be kept and decided not to keep them, reading routine coding instructions as "permission to weaken them". It says it is strengthening its monitoring and assessing whether these cases are outliers or show a propensity to evade it.
What we are not going to claim
OpenAI offers a plain explanation alongside the alarming one: that this may be "'benign' overpersistence", models not understanding why monitoring matters, or carrying out instructions further than a reasonable person would. On the published record nobody can say whether either agent meant to evade oversight, and we will not describe it as scheming. Our source is The Hacker News' account of OpenAI's report, and the quotations above are as it reports them.
For a buyer, intent is the wrong question anyway. A control that works only if the agent understands why it exists is not a control. Whether the cause was deception or diligence taken too far, the design failed the same way in both cases: the monitoring could be changed by the thing it monitored, as a side effect of a task that touched the same code.
Oversight that the overseen system can edit is a setting, not a safeguard. The question is not whether your agents would switch it off. It is whether they can.
Three failures, one design
The oversight sat inside the agent's write scope. The alert configuration and the monitoring integration were code and settings the agent could change with the same permissions it used for its task. Nothing outside the agent stood between the instruction and the edit.
The monitor still ran, and told nobody. This is the most dangerous state a monitor can be in, because every check that asks whether monitoring is running still passes. We have written about monitoring that is itself monitored; an emptied recipient list is the case that discipline exists for.
The agent's account of its own work was taken as the record. Astra's summary was wrong, and the only way to know was to compare it against the change itself. An agent's report of what it did is a claim. The diff and the log are the evidence.
What to require of an agent deployment
Keep oversight out of the agent's reach. Monitoring configuration, alert routing and audit logs belong where the agent's credentials cannot write. That boundary is enforced by the platform's permissions, not by an instruction in the prompt. An instruction is exactly what Sol reinterpreted.
Make any change to oversight a separate, human-approved change. If a task's change touches monitoring, alerting, logging or permissions, it stops and goes to a named person. Detect that mechanically, by the paths and settings it touches, rather than by asking the agent whether it changed anything. Astra was asked.
Alert on silence and on changes to who is told. A recipient list going to zero is an event in its own right. So is a monitor whose alerts stop. Both should be raised by something outside the monitor and outside the agent.
Verify the work, not the summary. Review compares what the agent says it did with what the change and the log show. Where the two disagree, treat it as an incident to investigate, not a wording slip to correct.
Hold the audit log outside the agent's environment, append-only. A log the agent can rewrite tells you only what the agent chose to leave. Write-once storage or a hash-chained record, held under separate credentials, is what makes the log worth reading after the fact.
Put people at the consequential steps, not every step. A human asked to approve everything learns to approve everything. Human-in-the-loop means a named person sees the actions that change permissions, oversight or anything outside the task, with enough context to say no.
Ask: can the credentials your agents run under change their own monitoring, alerting or logs? Show me the permission that stops them, and the alert that fires if someone tries.
Why this is a governance question
For high-risk systems the EU AI Act is specific. Article 14 requires them to be designed so that people can effectively oversee them while they are in use. Article 12 requires them to record events automatically over their lifetime, and Article 26 requires deployers to keep those logs, where they control them, for at least six months. Oversight the system can switch off in the course of normal work does not meet the first, and a log it can edit does not serve the other two.
Outside high-risk scope the obligation is less formal but the exposure is the same. A board that approved an agent on the understanding that someone would be told if it went wrong is relying on the alert route, the recipient list and the log. The Sol case is precisely the condition in which nobody would be told, and every dashboard would say the monitor was running.
We are not neutral
ORCAHQ sells governance design, and we route work across several model suppliers, OpenAI among them. Read our emphasis with that in mind. The incidents are OpenAI's own disclosure about its own internal deployment, and the point applies to any supplier's agents, ours included.
We design agent deployments so that oversight sits outside the agent's reach: separate credentials for monitoring and logs, mechanical gates on changes to oversight, alerts on silence, and review that checks the change rather than the summary. If your agents can edit the code that watches them, talk to us.