Permissions and revocation as an AI agent control point
Permissions and revocation is the control that limits what an agent may access and removes that access. On 16 Nov 2026, 5 of 16 vendors had a yes for this control point, 11 had partial and 0 had none found. Levels are from vendor documentation read that day, not lab tests.
Facts read on .
Permissions and revocation: every vendor
| Vendor | Level | Note | Evidence |
|---|---|---|---|
| CrowdStrike | Yes | Short-lived least-privilege access brokering and revocation when conditions change. | crowdstrike-31, crowdstrike-33 |
| Island | Yes | Per tool, per action and per actor limits through the MCP Gateway; just-in-time credentials. | island-13, island-15 |
| Microsoft | Yes | Conditional Access and Dynamic Permissions for agents. | microsoft-34, microsoft-55 |
| Noma Security | Yes | Tool and action level policy by identity. | noma-security-23, noma-security-17, noma-security-37 |
| Operant AI | Yes | Agent access enforcement and least-privilege MCP tool use. | operant-ai-19, operant-ai-21 |
| Aim Security | Partial | Over-privilege is spotted; revocation is not documented. | aim-security-47, aim-security-48 |
| Bay | Partial | Policy scoping and lock-down of reachable credentials; least-privilege revocation not specifically documented. | bay-31, bay-41, bay-42 |
| Check Point | Partial | Tool and MCP allow and deny lists; no least-privilege or revocation of agent credentials found. | check-point-27, check-point-21 |
| Cisco | Partial | Group-based policy and MCP server blocking; Duo IAM agent permissions planned. | cisco-16, cisco-31, cisco-40 |
| Lasso Security | Partial | Role-based permissions and overprivileged-agent flagging; revocation not documented. | lasso-security-24, lasso-security-34 |
| Palo Alto Networks | Partial | Identity governance for agents is documented; least-privilege and revocation details were not read. | palo-alto-networks-39, palo-alto-networks-40 |
| Pillar Security | Partial | Tool allow-lists per MCP server and RBAC for AI assets; no revocation found. | pillar-security-14, pillar-security-23 |
| SentinelOne (Prompt Security) | Partial | Least-privilege enforcement stated in general terms. | sentinelone-38 |
| Straiker | Partial | Overprivilege is identified and tool access can be revoked via the kill switch. | straiker-15, straiker-17 |
| WitnessAI | Partial | Organization-wide approved tool list; no credential least-privilege or revocation found. | witnessai-34, witnessai-32 |
| Zenity | Partial | Identity-aware policy and privilege visibility; revocation workflow not specifically documented in public pages. | zenity-35, zenity-36 |
What is permissions and revocation as a control point?
This control point sets least privilege for an agent and takes access back when the task ends or risk rises. It covers scopes on tokens, role ceilings and time limits on grants. AWS describes a permissions boundary as setting the maximum permissions an identity-based policy can grant to an IAM entity.
Sources: Agents are your newest insiders. Security must move at their speed..
Why it matters and where it stops
Microsoft says Dynamic Permissions let organization-wide policy set the outer boundary and scope authority to the task, with time limits and removal as the task changes. CrowdStrike says its identity provider brokers short-lived, tightly scoped access in place of standing credentials.
Permissions limit what an agent can reach, not what it chooses to do with that reach. They also depend on the agent using the brokered path: a credential stored in the agent's environment can bypass a broker.
Sources: CrowdStrike Announces Agentic Identity Provider; Permissions boundaries for IAM entities.
Who documents it
Yes on 16 Nov 2026: CrowdStrike, Island, Microsoft, Noma Security, Operant AI.
Partial: Aim Security, Bay, Check Point, Cisco, Lasso Security, Palo Alto Networks, Pillar Security, SentinelOne (Prompt Security), Straiker, WitnessAI, Zenity.
None found: none.