Unrelenting entrepreneur Founder of @permit_io, @Rookoutlabs, @Reactful Permissions for the AI Era (https://nitter.cf/t.co/3jbOiNRcAl)
Joined September 2013
- Tweets1.1K
- Following156
- Followers1K
- Likes2.2K
WorkOS shipped Fine-Grained Authorization in February: RBAC plus resource hierarchy, with permissions flowing down from workspace to project.
If you're evaluating it, here's an honest comparison with Permit.io, including when WorkOS is the better pick 🧵
Where Permit differs: policy is generated as Rego or Cedar you can keep in Git, and decisions can run in Edge PDPs inside your network, working from cached policy if the control plane is unreachable.
Also: Zanzibar-style relationship tuples for deep sharing graphs, plus Permit Elements so your customers' admins can manage access themselves.
Full side-by-side by Ze'ev Manilovich:
permit.io/blog/workos-fga-al…
APP 1.7 homework for engineering, using the four authorization pillars. Applies from 10 Dec 2026 to Australian orgs whose programs use personal info in significant decisions.
Granularity: map kinds of PI to resources, actions and tenants, and tag each path as solely automated or assisted. Realtime: list the changes that trigger a re-review and make sure revokes reach enforcement.
Audit: decision logs privacy and legal can use, denies included. 10-item checklist here (not legal advice):
permit.io/blog/four-pillars-…
APP 1.7 to 1.9 start in Australia on 10 December 2026. If a program uses personal info to make or substantially assist a significant decision, the privacy policy has to disclose the kinds of PI and kinds of decisions involved.
The risk to watch is drift. A new model, data source or auto-approval step can quietly make an accurate policy statement stale.
Engineering side plus a 90/60/30-day countdown (not legal advice):
permit.io/blog/au-app17-apra…
NIST NCCoE's concept paper on AI agent identity and authorization names four areas: identification, authorization, access delegation, logging. We mapped each one to a concrete control.
Delegation: an explicit human-to-agent grant with scope, duration and revocation.
Logging: which human, which agent, which tool, which data, allow or deny, and why.
The paper is a concept document seeking input, so use this for design reviews. It won't certify anything. Cheat-sheet:
permit.io/blog/nist-nccoe-au…
EN 18286 is the first harmonised standard under the EU AI Act, now in the Official Journal.
It covers QMS for high-risk AI. For agent builders, the question it raises is whether you can prove what your agent was allowed to do and who approved it.
permit.io/blog/en-18286-qms-…
Most agents in prod borrow a shared API key or the user's session. Whatever that key reaches becomes the blast radius, and the audit log says the human did everything.
For MCP, OAuth covers whether the client can connect. Whether this agent can call this tool on this resource right now is a separate check. A gateway PEP in front of MCP servers can do it without rewriting each server.
NIST NCCoE's concept paper frames the same four areas: identification, authorization, delegation, logging. Full post:
permit.io/blog/agents-as-pri…
