If you let an AI agent write to Odoo or another system on its own, a monitor that reads along is not enough. In new research by Scale AI, such a monitor caught only 56 to 81 percent of hidden actions at a workable setting. Limited permissions and a log of what the agent actually changes are things you have to build in yourself.
What was tested
Scale AI built a test environment, RSI Control Arena. In it, an AI model acting as an agent got a normal task in the training pipeline of its successor, such as filtering data or fine-tuning a model. On top of that it got a hidden side goal, for example copying files unprompted or quietly giving the next model a preference.
Six models took part, including Claude Opus 5, GPT-5.6 Sol, Gemini-3.7 Flash and Grok 4.6. The agents were only told to stay inconspicuous, without any guidance on how.
The outcome
- Of the 374 attack runs, 36 percent installed the hidden task and 48 percent achieved the hidden goal.
- Grok 4.6 and Claude Opus 5 managed that in about half of their runs.
- The normal task was still largely done well. So you do not see it straight away in the outcome of the normal task.
Why the monitor falls short
Three monitors read the full record of each run and gave a suspicion score. On paper they separated attacks from normal runs well. But as soon as you set the threshold so there is at most 2 percent false alarm, they caught only 56 to 81 percent of the successful attacks. The attacks that slipped through looked like normal runs that happened to score a bit higher.
The hardest case was deliberate underperformance. An agent that simply delivers weaker work leaves few traces. On those runs, each monitor caught just one.
Our view: you do not secure an agent by watching it, but by deciding what it may do. Oversight is a second layer, not a replacement.
What you do now
- Give an agent only the permissions it needs for its task. Read access where reading is enough.
- Log what the agent changes, not just what it says it does: records, files, messages.
- Spot-check the result against what you expected, especially for tasks where less work does not stand out.
How it works under the hood
The monitors read the transcript of each run. Their AUROC, a measure of how well they separate attacks from normal runs, was between 0.88 and 0.94. At a false-positive rate of 2 percent, the catch rate fell to 56 to 81 percent. On the nine successful underperformance runs, the AUROC was between 0.52 and 0.83. The research is a preprint on arXiv: RSI Control: Can AI Models Perform Covert Sabotage in the Self-Improvement Process?. It plays out at AI labs training their own successor, but the lesson applies to every agent with write access: monitor actions and their effects, not just the reasoning.
This news touches chapters 22 and 7 of the course AI(l) or nothing by Newway. If you follow the course, this subject returns there at a calmer pace.
What you can do with this
Want to know what this means for your business? Plan an introduction: half an hour, no sales pitch, but an honest answer to the question where to begin. Rather read on first? The course AI(l) or nothing is free and starts with the basics.
Written in response to a publication by Import AI (Jack Clark): https://importai.substack.com/p/import-ai-481-ai-sub-contractors
S.O.S