Dev.to Security 🔐 Cybersecurity 👁 0 📖 2 min read

Why "deny or allow" is the wrong model for MCP tool permissions: a verdict table with reason codes

If you build agents on MCP, you have met this loop: a tool call should not run, the framework raises a permission error, the operator retries, and nobody can later explain what the system believed at the moment of decisi

If you build agents on MCP, you have met this loop: a tool call should not run, the framework raises a permission error, the operator retries, and nobody can later explain what the system believed at the moment of decision.

We spent the last months turning that loop into a deterministic decision table — published, replayable, and deliberately free of an LLM in the gatekeeping path. Three design choices surprised even us:

1. ESCALATE is a verdict, not a failure.
Irreversible actions route to a human gate. That is not a denial with extra steps — folding it into a permission error is exactly how approval fatigue starts. The operator starts rubber-stamping because the system cannot say why it paused.

2. Attempt level, not action level.
Two attempts can share one logical action id (hello, lost-response retry). The decision table keys verdicts at the attempt level; the action id aggregates them for audit display, never for the decision itself. This matches the shape the crewAI community converged on in issue #6221 independently.

3. unknown never folds.
If the evidence is partial, the verdict is unknown, with its own code. It does not round to PASS and it does not round to DENY. Audit consumers deserve the honest column.

The table anchors 576 parameter-risk decisions across mutating and read-only axes, and every verdict ships with the reason code so a downstream system can distinguish:

  • TOOL_DENIED_BY_POLICY — policy said no
  • VERDICT_ESCALATE — parked for a human gate
  • INTEGRITY_TAMPERED_EVIDENCE — the record itself failed integrity; trust nothing downstream of it
  • SIDE_EFFECT_REPLAY_RISK — retried mutating call, judged at its own attempt

It also runs as a GitHub App (Correctover Admission Verifier) so verdicts land in Checks with the same codes.

The full dataset is CC0, the code is Apache-2.0, no CLA, no telemetry:
https://github.com/DSHCorrectover/correctover-recovery-admission

If your framework or platform needs an independent look at its admission semantics — we do bounded review engagements, fixed price, every finding ships with a replay pack. [email protected].

📰 Read the original article on Dev.to Security

Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.