Some colleagues trust an agent with root access across production because it has worked well so far and its actions are logged. I want to keep using the tool, but with better limits. How can I explain the risks without turning it into a confrontation?
How do I explain the risk of giving an AI agent root access to production?
Answers
Start with a concrete task rather than arguing about whether the model is trustworthy. If it needs to inspect a service, give it a read-only identity scoped to that service. If it needs to make a change, require approval and grant only the access needed for that change, ideally for a limited time.
Logs tell you what happened after an action. Permissions can stop the action in the first place. Also, if the agent has root on the machine holding its controls or logs, it may be able to alter them. Keep audit records and backups outside the agent's control, and test changes in staging first.
Reply to this answer
Prompt injection is a different problem from an ordinary bad command. The agent may read a log entry, ticket, or web page containing instructions that conflict with its actual job. If it treats that content as an instruction and has broad permissions, the consequences can reach well beyond the thing it was asked to inspect.
Try a harmless exercise in a sandbox with a planted instruction in test data. Check what actions the agent attempts and what its permissions actually allow. That makes the boundary concrete without risking production.
Reply to this answer
Bring security or whoever owns production risk into the conversation. Present the access, the possible consequences, and a few guardrail options, rather than making it about colleagues being careless. Record who approves the setup and who is responsible for agent actions.
I'd also ask how credentials and customer data are kept out of requests to the model provider. Root access can make that data available to the agent, so the team's privacy and contractual requirements need a look too.
_d68~plivaro.5:
Access to sensitive data can still be necessary for some operational work. The useful question is who or what needs it, for which task, and under what controls.
Reply to this answer
Don't use production as the experiment. A clean run so far doesn't prove the next one will be clean, especially if the provider changes the model. Test representative tasks and failure cases, including what actions the agent tries to take, whenever the model or harness changes.
And check recovery separately. Can the team restore a known good state if something goes wrong? Backups only help if the agent can't modify them and the restore process works.
Reply to this answer
Add your answer
Answers are reviewed before publication.
q.836+brasswicket:
That gives me a useful way to frame it around the task instead of whether anyone trusts the model.