| Level | Name | What the AI may do | Used by |
|---|---|---|---|
| L0 | Suggest | Recommends; a person acts | Document intake, behaviour model, STR summariser |
| L1 | Act with approval | Prepares the action; a person approves it | Copilot, fix agent, correlation agent |
| L2 | Act within guardrails | Takes low-risk actions inside limits and reports afterwards | Triage agent (closes noise with reasons) |
| L3 | Autonomous with review | Acts; people review a sample | Not used |
| L4 | Coordinating agents | Agents delegate to other agents | Research · after 2030 |
| L5 | Self-modifying | Changes its own policies | Not permitted |
| Agent | Tools allowed | Limits | Kill switch |
|---|---|---|---|
| Security Copilot | navigate, lookup_ip, block_ip, set_geo_rule, raise_incident | Blocks expire in ≤ 30 days; no Whistleblower rules | Settings › Break-glass |
| Triage agent | close_alert, escalate_alert | Can close only below P = 0.15 | SOC lead |
| Fix agent | open_pull_request | Cannot merge | Engineering lead |
| Correlation agent | group_alerts, propose_fix | Read-only on infrastructure | SOC lead |