Skip to main content
Security Analysis

Security Model & Threat Assumptions.

Explicit threat model for Attested Governance Artifacts. What we defend against, what we do not, and the assumptions that underpin our guarantees.

Security Goals

What AGA Proves.

AGA provides verifiable decision integrity, tamper-evident governance records, and portable audit evidence.

It generates cryptographic proof of every governance decision it recorded: each one signed, chained, and independently verifiable. It does not prove the subject took no ungoverned action; that is a deployment property.

Adversary Model

Assumed Attacker Capabilities.

The adversary model assumes an attacker with full access to the local database and network but not the cryptographic signing keys or gateway runtime.

AGA provides structural integrity guarantees independent of payload confidentiality. The system is designed so that even with full visibility into stored data and network traffic, an adversary cannot forge valid receipts or tamper with the governance chain without detection.

Threats Addressed

What AGA Defends Against.

Binary modification

Hash comparison against sealed baseline

Configuration drift

Continuous measurement at policy-defined cadence

Unauthorized dependency changes

Checksum validation of loaded modules

Runtime policy bypass

gateway as mandatory execution boundary

Audit log tampering

Hash-linked receipt chain

Evidence fabrication

Merkle checkpoint anchoring to append-only storage

Replay attacks

Timestamped receipts with sequence numbering

Privilege escalation

Two-process separation; subject holds no signing keys

Supply chain model substitution

Sealed reference hash binds identity at build time

Behavioral drift in tool-call patterns

Statistical analysis over sliding windows, producing signed BEHAVIORAL_DRIFT receipts

Policy bypass via indirect execution

Phantom execution with forensic capture detects circumvention attempts

Threats Not Addressed

Explicit Exclusions.

  • Kernel-level or hypervisor compromise of the gateway host
  • Hardware microcode or firmware attacks
  • Adversarial prompt manipulation at the model level (inference-time semantic defenses are complementary and out of scope)
  • Side-channel attacks on the signing key
  • Social engineering of the policy issuer

Note: AGA addresses behavioral drift at the tool-call level through statistical analysis over sliding windows. Model-level semantic defenses remain complementary and out of scope.

Trust Assumptions

Foundation of Guarantees.

gateway Integrity

The gateway must execute within a trusted boundary. If compromised, receipt integrity cannot be guaranteed. For highest assurance, the gateway can ingest TEE attestations where present, though this is not required for standard operation.

Key Custody

The policy issuer's signing key must be protected. If the signing key is compromised, forged artifacts become possible. Key rotation is supported through the continuity chain mechanism.

Verification Guarantees

Assurance Levels.

LevelStepsGuarantees
OfflineSix offline checks
  • Artifact authenticity
  • Receipt chain integrity
  • Merkle proof validation
  • Signed checkpoint
OnlineOptional anchor
  • Checkpoint anchor validation against append-only storage
FullOffline checks + anchor
  • End-to-end governance verification

Explore the System.

See how these security properties are implemented in practice across the governance architecture.

MCP GOVERNANCE

Zero-TOCTOU Property for MCP Agents

For MCP-governed AI agents, with the proxy deployed as the sole tool channel, the time-of-check-to-time-of-use window is zero by construction. The gateway is the communication channel. Every tool call passes through the gateway synchronously. The agent cannot invoke a tool without the gateway evaluating it first.

If the gateway is the only path to external resources (applied by network isolation), and every tool call is synchronously evaluated before forwarding, then unauthorized external actions are recorded and, when the boundary is wired as the sole egress path, refused. Internal computation is unconstrained, but internal computation without external action causes no harm. The moment the agent acts, it hits the gateway.

This property is specific to MCP-governed agents where the proxy is the sole communication path. For container and binary subjects, AGA uses periodic measurement with configurable cadence (100ms-5s) and event-driven eBPF probes on the roadmap.

PHANTOM EXECUTION

Phantom Mode Limitations

When a quarantine decision is sealed, the artifact specifies phantom mode; in the reference orchestrator the agent receives synthetic responses while actions are captured. Severing real effects is wired per deployment. This captures adversarial intent without alerting the agent.

A sufficiently capable agent could potentially detect phantom mode through timing analysis, response content patterns, or side channels. AGA mitigates this in the reference orchestrator by profiling real response latencies during normal operation and replaying from the recorded distribution during phantom mode. The JSON-RPC envelope structure matches exactly.

The agent would need to specifically allocate reasoning capacity to phantom detection, which requires knowing phantom mode exists. The purpose is to capture intent during the observation window, not to maintain indefinite deception against a determined detection effort.

RESOLUTION STATUS

Logical Problem Resolution

Analysis of the AGA architecture identified 12 non-trivial logical problems. Ten are addressable with engineering; two are irreducible residuals present in any cryptographic governance system.

ProblemCategoryStatusMitigation
gateway trustFundamentalMitigatedThreshold signing (roadmap); TEE attestation as measurement input (deployer-provided)
Key compromiseFundamentalMitigatedKey rotation and PQC hybrid (implemented); HSM isolation (roadmap)
Policy correctnessFundamentalIrreduciblePolicy provenance chain, coverage metrics
TimestampsArchitecturalMitigatedRoughtime IETF integration, monotonicity verification
Identity bindingArchitecturalMitigatedSPIFFE self-signed X.509, certificate chain binding
TOCTOU windowArchitecturalSolvedMCP proxy is synchronous by construction
Proxy bypassArchitecturalAddressed (deployment)Mandatory network isolation at infra level
Phantom detectionArchitecturalMitigatedTiming profiles, envelope matching
Recursive governanceArchitecturalRoadmapTEE-hosted gateway with attestation chain
Policy rotationOperationalSolvedPOLICY_ROTATION receipt type, zero-gap protocol
Performance at scaleOperationalMitigatedReceipt batching, horizontal scaling
Legal admissibilityOperationalIrreducibleStandard primitives, expert witness readiness