
The flag marks the client as an agent. It adds no scopes and doesn't decide which tool this agent may call for this user right now. That's still per-tool policy and an audit log. Full walkthrough: io.permit.io/gws-mcp-x

Unrelenting entrepreneur Founder of @permit_io, @Rookoutlabs, @Reactful Permissions for the AI Era (Agent.security)

The flag marks the client as an agent. It adds no scopes and doesn't decide which tool this agent may call for this user right now. That's still per-tool policy and an audit log. Full walkthrough: io.permit.io/gws-mcp-x




ABAC is the same middleware with richer data (department, invoice amount), and the limit lives in policy, not code. I ran this on Express 5.3.0, the permitio Node SDK 2.7.7 and @nodejs 24 against a stub PDP, with five passing tests: io.permit.io/express-authz-x

Keep the middleware shape and move the decision to a PDP. With @permit_io that's permit.check(user, action, resource). The JWT identifies the user. Roles live in Permit per tenant. A revocation applies on the next check once it reaches the PDP. If the PDP is down, return 503.



Authorization in @expressjs is one decision per request: can this user perform this action on this resource? Express gives you the middleware hook. The decision is yours, and I walked through four steps with running code 🧵

Both come down to control-plane authorization: who can create, rename, own and change adapters and tools, and what the gateway does with the values they supply. I put a checklist in the post: io.permit.io/mcp-gw-sec-x


CVE-2026-109492 is a CWE-639 authorization bypass. @Microsoft scored it CVSS 8.8. User-chosen names became Kubernetes object names in one namespace, so a user who could create adapters or tools could take over another workload or the built-in tool gateway. Fixed in 1.0.0.

Two CVEs in MCP gateways, both about who can change the gateway itself. I wrote up CVE-2026-109492 in the open-source @Microsoft MCP Gateway and CVE-2026-125926 in mcp-gateway-registry, disclosed by @awscloud. 🧵

After the exchange you know who the user is. Whether their agent should call delete_customer_data after reading untrusted content gets decided per tool call. @zivxx wrote it up with a full validation checklist (code samples are schematic): io.permit.io/mcp-ema-x

