Why this is the question for a lean team
A large SOC can afford to approve every action by hand. A team of two or three cannot, which is the reason it is buying an AI SOC. But handing over actions without limits is not an option either. The useful middle is a set of rules that say which actions may run alone, under what conditions, and how that permission is withdrawn. Reversibility, meaning how easily an action can be undone, is the most practical of those rules.
This note compares how each vendor's public pages describe that middle. It reflects pages read on 27 September, 11 October and 13 October 2026.
Named levels with action checks: Mate Security
Mate describes three autonomy levels: Recommend only, Supervised and Auto-run. It says each action is scored on Confidence, Precision and Reversibility, which it calls CPR, and that these checks are enforced outside the agent loop. A category drops back to Supervised automatically when its verdicts conflict with analyst reviews. Of the ten, this is the only public description that scores each action on reversibility as an explicit check and includes an automatic way back.
Pre-authorised per action and per connector: Artemis Security
Artemis Security describes acting alone only where you have pre-authorised it, with approval set per action and per connector. Its containment is impact-checked and described as reversible by design, and its audit trail records who, what and why. A way for autonomy to come back automatically is not described.
Scoped actions with optional sign-off: Prophet Security
Prophet describes scoped Agent Actions that run autonomously or with your sign-off. Scope is itself a limit on blast radius: an action defined narrowly can only do so much. Prophet also sells Watchtower, a 24x7x365 human-in-the-loop service, which puts people on the loop without the customer staffing them.
Approval gates and an audit trail: 7AI
7AI describes response that is guided or automatic, with humans on the loop, behind approval gates and a full audit trail. Its examples, revoking a session, resetting a password, isolating a host, are mostly reversible actions, which is a sensible place to start.
Approval gates and Slack confirmations: Exaforce
Exaforce describes Exabots that execute response behind approval gates, with full audit trails and Slack confirmations from users or managers, and actions described as approved and reversible. Named autonomy levels per category, and a way to pull autonomy back, are not described.
Staged progression: Conifers
Conifers describes starting with a human in the loop, moving to a human on the loop, and expanding autonomy after validation. Its May 2026 release added remediation within set limits. The stages are clear; the per-action rules behind them are described in less detail.
Autonomy through workflows: Torq
Torq runs response through Hyperautomation workflows, deterministic or agentic, with governance and audit logs, and states that over 90% of cases are remediated completely autonomously; that figure is Torq's. Approval steps are something you build into workflows rather than a named mode on the pages we read.
Response in your SOAR: Dropzone AI
Dropzone AI hands response to the SOAR you connect. Your existing approvals stay where they are, which is reassuring if they work, but the control over what runs alone is in another product.
Automatic or manual: Intezer
Intezer describes remediation plans executed automatically or manually, with human teams reviewing outcomes. Graded levels between those two are not described.
Recommendations only: Vega
Vega does not state that it executes response. Verdicts come with prioritised next actions, such as isolating a host or rotating keys, for analysts to carry out, so there is no autonomy to set.
What to take from the comparison
Three questions separate the descriptions:
- Can you set autonomy per action, not only per product? Mate, Artemis, Prophet and 7AI describe this most directly.
- Is reversibility, or blast radius, an explicit input to whether an action runs? Mate names it as a scored check, and Artemis describes its containment as reversible by design; for the others, you apply it yourself through scope and approvals.
- Does autonomy come back automatically when something drifts? Mate describes an automatic fallback; for the others, plan a manual routine.
These are the questions behind our autonomy criterion, worth 20% of the hand-off score. Public pages describe intent. In a pilot, ask each vendor to show you the setting screen, change one action from unattended to sign-off, and show the audit record of the change.
A starting list of reversible actions
For a team setting rules for the first time, these actions are usually quick to undo and narrow in effect: quarantining a reported email, revoking an active session, adding a temporary block on a single file hash, and isolating one endpoint that can be released with a click. Actions that are hard to undo or wide in effect, such as deleting mail across all mailboxes, disabling an executive's account or blocking a domain the business uses, belong on sign-off until your own evidence says otherwise.