AI SOC autonomy levels explained: from recommend only to auto-run

Vendors name them differently, but most describe three. Mate Security's are the clearest example: Recommend only; Supervised, where low-risk steps run and risky changes need approval; and Auto-run for trusted categories under guardrails and audit. Conifers describes the same progression in other words: start with a human in the loop, move to a human on the loop, and expand autonomy after validation.

Human in the loop means a person approves an action before it runs. Human on the loop means actions run automatically while a person monitors and can intervene. The first is slower and safer; the second is faster and depends on good monitoring and a quick way to stop the agent.

• Approval gates on specific actions: Prophet Security offers scoped Agent Actions, run autonomously or with your sign-off; 7AI describes approval gates and a full audit trail; Palo Alto says Cortex AgentiX can require approval for high-impact actions and binds agents to the same roles and permissions as analysts.
• Guardrails outside the model: Mate says its guardrails are enforced outside the model's reasoning loop.
• Automatic fallback: Mate says it rolls back to Supervised mode when behaviour drifts.

Per alert category, not for the whole platform. Start every category at the lowest level, raise it only when verdicts have matched analyst decisions for a sustained period, and keep high-impact actions such as disabling accounts or isolating servers behind approval for longer. Our proof of value note describes how to gather that evidence.

Human-in-the-loop controls carry 15% of the weighted score in our method. The highest score on this criterion is 9, held by Mate Security and Artemis Security. The criterion rewards approval gates, autonomy levels set per category, scoped actions, an audit trail and a quick way to pull autonomy back.