Sharing what I'm building: ZIFFER, a deterministic authorization layer for AI agents.

Evaluationby Yacine "ZIFFER"2026-09-16

Sharing what I'm building: ZIFFER, a deterministic authorization layer for AI agents. The model proposes an action, ZIFFER decides at the tool-execution boundary. Policy-as-code in git, signed decisions, full audit trail. Model-agnostic, so it sits next to any LLM, including Jev once I'm off the waitlist. Repo: https://github.com/yacine-kellib/agent-control-plane (Apache-2.0, open spec, Dafny-verified invariants, 20k-case test harness) Hosted version: https://ziffer.io Curious how typed System One outputs pair with an external policy check. Feedback and holes welcome.

Links

yacine-kellib/agent-control-planegithub.comA credential is not authorisation. ACP moves the authorise/refuse decision for AI agent actions outside the model, where prompt injection cannot reach it. Specification, Dafny proofs, and a referen...

ZIFFER: agent authorization. An injected agent can’t act.ziffer.ioYour agents will get injected. With ZIFFER they hold no credential, so they cannot act. Every action is a proposal your policy grants or refuses.

open source discussion

More in Evaluation

Return probabilities, not labels — 80/10/10 beats "orange"I’m building Aurapunk, an open-source multi-agent IDE.Tested jev against deepseek v4.1 flash on norwegian text.