# The Agent Unplugged the Alarm

> On a test floor the safety interlock is wired outside the unit under test. OpenAI just reported two coding agents that weakened their own misalignment monitoring. Agents need the same rule.

- Canonical URL: https://webofmike.com/agent-unplugged-the-alarm/
- Author: Mike Moore (https://webofmike.com/about/)
- Published: 2026-11-02
- Last modified: 2026-11-15
- Tags: AI Agents, Security, Thoughts
- Cite as: Mike Moore, "The Agent Unplugged the Alarm", Web of Mike (webofmike.com), 2026-11-02. https://webofmike.com/agent-unplugged-the-alarm/


On a test floor, the safety interlock is never wired through the unit under test. It sits on the fixture. The door opens, the power drops, and nothing the DUT does can change that. If the board could jumper its own interlock, you wouldn't have an interlock. You'd have a suggestion.

The one on our high-voltage station had a single strip of red electrical tape folded into a flag. It was there before me and it was there when I left.

OpenAI just published a report that shows why agents need the same rule.

## What OpenAI reported

The report is [Suppressing misalignment monitoring alerts and disabling misalignment monitoring](https://alignment.openai.com/misalignment-reports/suppressing-misalignment-monitoring-alerts-and-disabling-misalignment-monitoring). Two internal coding tasks, two agents, and in both cases the agent weakened the monitoring that watches agents while doing ordinary software work.

**October 3, the alerts.** A researcher asked a GPT-6.1 Sol agent to build eval tasks from employee feedback in Slack, and told it "don't send any messages." Much later, the pipeline stopped after its monitor flagged a proposed repair and couldn't save the finding. The researcher asked the agent to point the monitoring results somewhere it could access and rerun.

The agent first read "don't send any messages" as being about Slack outreach, not monitoring. A subagent it delegated to considered setting `no_notify` and decided against recommending it, because disabling required alerts would be unsafe. Then the parent agent changed its mind and told the subagent to suppress notifications to honor the user's no-messages request. The patch emptied the ping and page recipients, set `no_notify` to true, and made suppression a precondition for writing a result. The monitor still ran and saved its reports. The humans just stopped getting told.

**October 6, the monitor itself.** A user asked a GPT-6 Astra agent to copy the serving path from a reference implementation. Earlier, it had been told not to disable the monitor "even temporarily." The agent delegated the migration and told the subagent to preserve monitoring "as appropriate to exact PR." The subagent read that as permission to drop the monitor, since the reference didn't have one.

Then the part that got my attention. The agent replaced a regression test that was meant to preserve monitoring with one that required monitoring to be gone:

```
assert "monitor_ctx" not in saved["context"]
```

When the user asked whether the path still had L1 monitoring and whether it had been removed, the agent said it "did not remove an enabled monitor from it."

## Why it slipped through

OpenAI's own summary is the important line: in both cases, the agents considered whether monitoring safeguards should be preserved, but interpreted routine coding instructions as permission to weaken them.

Neither incident was classified as P0. OpenAI pages Security immediately only for P0, and neither trajectory was killed automatically.

The report also says OpenAI has seen several cases where an agent launched a deployment with misalignment monitoring disabled. The agent itself stayed monitored. The thing it launched didn't.

That's the DUT jumpering the interlock on the next station down the line.

## Where the interlock lives

The useful detail is where monitoring was hard to turn off and where it wasn't. Per the report, OpenAI expects it to be difficult to disable for Codex workloads, because monitoring runs in the service that handles model requests, outside the agent's direct control. It was easier for evaluations, partly because monitoring isn't currently required for weaker Luna-class models, so an Astra eval could be misconfigured to use the bypass meant for them.

Same company, same agents. One path had the interlock on the fixture. The other had it as a config flag the agent could edit.

OpenAI's fixes:

1. Auto-Review, already required for internal Codex use, now has specific instructions to detect and block attempts to disable misalignment monitoring.
2. They're working toward requiring a valid safety case for every category of workload, with workloads that lack one, or don't meet it, terminated automatically.

They also say plainly that they don't yet have a satisfying explanation for why models do this. It might be benign overpersistence. It might be something more specific. They're building evals to find out.

## What I'd take from it

You don't need to know why the DUT shorted the interlock to know the interlock was in the wrong place.

- **Monitoring runs outside the agent's reach.** In the request path, at the gateway, in a service the agent has no credentials for. Not in a config file inside the repo it's editing.
- **Tests that guard safety controls are not ordinary tests.** If an agent can rewrite the test that checks the monitor, the test checks nothing. I wrote about this in [your tests are lying to your agent](/tests-lying-to-your-agent/).
- **Suppressing an alert is a change to the alarm system.** "Don't send messages" from a user should never be enough to silence a page to Security. That decision needs its own approval, from someone who owns the alarm.
- **Whatever an agent launches inherits the monitoring requirement.** The child workload doesn't get to start without it.
- **Ask the system, not the agent.** The Astra agent answered a direct question wrong. If you want to know whether monitoring is on, check the monitor's own record. Same idea as [keeping the evidence when you stop an agent](/stopping-an-agent-without-losing-evidence/).

And when the alarm goes off and the page goes nowhere, you end up with [the pager going off and nobody waking up](/pager-goes-off-nobody-wakes-up/).

Wire the interlock to the fixture. Then let the agent do whatever it wants inside the box.

## Sources

- [Suppressing misalignment monitoring alerts and disabling misalignment monitoring](https://alignment.openai.com/misalignment-reports/suppressing-misalignment-monitoring-alerts-and-disabling-misalignment-monitoring), OpenAI Alignment, report updated October 20, 2026.

