Saltar al contenido principal

warden

Since v2.43.0 · Used inside a declaration

Grammar​

warden(<target>) within <Scope> {
    // find_exploits() -> list[Vulnerability]  (attested findings only)
    // fortify(findings)                         (proposes hardening)
    <body statements>
}

warden is the authorization-native adversarial security-analysis primitive: it audits a target system to find (and propose how to close) vulnerabilities. Distinct from shield — shield is a passive I/O firewall of the agent itself (prompt-injection/PII); warden is an active auditor of a target, behind an authorization gate.

warden reframes the model as an adversary (Popperian falsification): it assumes the system is compromised and searches abductively for the evidence that would prove it — but only over authorized evidence, and it emits only attested Vulnerability findings (a re-checkable witness: the input + trace + violated contract), never fuzzy prose.

Surface​

warden is a flow-body block (like quant). The within <Scope> clause is mandatory — a scopeless warden cannot be written.

scope InternalAudit {
    targets: [ "svc://payments-core" ]
    depth: static_artifact
    approver: requires "security.lead"
}

flow Audit() -> Unit {
    warden(payments_core) within InternalAudit {
        step Analyse { ask: "enumerate contract violations" }
    }
}

The five governance guarantees​

  1. Signed scope mandatory — warden(t) within <Scope>; with no resolvable scope the program does not compile (axon-T887, fail-closed by grammar).
  2. Authorized evidence only — the analysis ingests operator-provided artifacts within the scope's allowlist; live memory/network capture is the most-restricted, enterprise-only depth, never a default.
  3. Attested findings — every Vulnerability carries a witness; an un-witnessed finding is rejected and retried via immune. Precision over recall.
  4. Audited + fail-closed — every analysis is hash-chain audited (scope + evidence digests, never raw bytes); an out-of-allowlist or unapproved analysis is refused.
  5. fortify proposes, a human applies — hardening is a typed, reviewable diff per finding; applying it passes through an approval gate. warden finds; an authorized operator decides.

What this primitive is NOT​

  • Not shield. shield defends the agent's I/O; warden audits a target. Orthogonal axes.
  • Not an autonomous exploitation tool. It is a defensive / authorized-testing instrument by construction — no unscoped, un-audited mode exists. It finds and proposes; it does not weaponize.
  • Not a claim to out-hack humans. Its superiority is governance + verifiability (authorization-native, witnessed findings), not raw offensive capability (the no_unwitnessed_advantage discipline).

See also​

  • axon://primitives/scope — the mandatory authorization envelope.
  • axon://primitives/shield — the passive runtime I/O firewall of the agent (a different security layer).
  • axon://primitives/mandate — the approval gate fortify applies through.