Agent Firewall

Security control plane

version
depth policy
decision source
Local inspection console. Read-only view of the existing authorization implementation. It makes no security decisions of its own — every verdict below is returned by the SDK. Cryptographic material is withheld. No authentication; bind to loopback only.

North Star pipeline

Ordered gates read from _authorization_gate_phases() — not a hardcoded diagram.

passed denied not reached

    Decision

    awaiting request

    Delegation authority

    Risk & security posture

    Capabilities

    Signatures and public keys are withheld by the projection layer.

    Lifecycle & evidence

    Append-only event record, newest first.

    Authorization trace

    Canonical SecurityDecision as returned.

    Agent identity & supply chain v2.0

    Agent identities with full lifecycle, verifiable security passports, and supply-chain provenance. Identity proves who; the authorization pipeline alone decides what.

    Identities

    Lifecycle, key fingerprints, parent/child. Mutations go through the control plane.

    Security passport

    Deterministic, signed summary of identity + posture.

    Supply-chain provenance

    Models, tools, MCP servers, skills, plugins, packages, adapters, configuration, policies. A name is never trust.

    Security operations v1.9

    Cross-agent security intelligence over verified artifacts: what agents can do, what they are doing, what could happen next, and how to respond safely.

    Active agents

    Reachable capability, tools, resources.

    Suspicious behavior

    Deterministic, evidence-backed detections.

    Correlation

    Artifacts grouped by shared identifiers.

    Sensitive resources

    Attack-path summary over recorded reach.

    Attack paths

    Recorded paths from an agent to a target, with break-path suggestions. Reachable is not exploitable.

    Scenario simulator

    What happens if this agent is compromised? Isolated workspaces only; never touches live state.

    Security flight recorder v1.8

    Portable, cryptographically verifiable security history. The artifact records what happened and why; verification, replay, and the graph are derived from recorded events only.

    Agent security timeline

    Every event is inspectable: from timeline to event to decision to authority chain to evidence.

    Security trajectory

    How posture changed, and the evidence for every transition.

    Security relationship graph

    Derived from recorded authority and decisions. Why could this agent do this? What could it reach?

    Containment

    Explicit state transitions, audited. Mutations go through the control plane.

    Replay laboratory

    Ask what would have happened under a different policy. Replay runs the real authorization pipeline in throwaway workspaces.

    v2.1 Autonomous Defense v2.1

    Defense mesh, a2a zero trust, attack graph, digital twin, evidence graph, immune system, and research lab - advisory and analytic layers above the authorization pipeline.

    Live defense state (mesh).

    Agent-to-agent trust relationships.

    Attack-graph summary (reachability is not exploitability).

    Digital twin counterfactuals.

    Evidence graph (signed, tamper-evident).

    Immune system: OBSERVE..VERIFY loop. Model output is advisory only; execution requires a policy rule.