Security Analysis

AIVSS Risk Mapping

This page compares the controls in APS with the ten risks of the OWASP AI Vulnerability Scoring System and explains their limits. It is an assessment by the APS project. It is not an OWASP assessment, a certification or an AIVSS score.

How to read the three groups. The first group lists risks where APS has cryptographic or deterministic enforcement with adversarial tests. The second lists risks where APS offers detection or structural mitigation and no hard prevention. The third lists risks that APS names and barely mitigates. The APS developers wrote all the adversarial tests, and no independent red team validated them. The entries describe the SDK as analyzed in the paper and have not been rechecked against SDK 8.0.1.

Risks with relevant APS controls

#1 Identity Spoofing

Ed25519 key pairs bind identity to cryptographic material. Passport signatures are verified on every protocol interaction. An attacker cannot impersonate another agent without possessing its private key. The deployment still has to protect the keys and decide which identities it trusts.

INV-1Layer 1: IdentityAdversarial: S1

#2 Access Control Violation

Scope monotonic narrowing enforced at every delegation step. The deterministic gate checks scope on every action. Sub-delegations cannot exceed parent scope. The erratum of 2026-08-17 concerns the older Delegation records. The AuthorityDelegationV1 verifier in SDK 8.0.1 requires all seven limits and rejects a wider spend limit or depth.

INV-2INV-3Layer 1: DelegationLayer 5: PolicyAdversarial: S2

#3 Cascading Failures

Revoking a delegation cascades to all descendants. Revocation is irreversible. Revoking the delegation of a compromised intermediate agent removes the authority of its whole sub-tree once the verifier has the new state. It does not undo an action that already ran.

INV-4INV-5Layer 1: DelegationAdversarial: S5

#4 Untraceable Agents

A delegation chain identifies its root. Showing which human stands behind that root needs separate evidence. The beneficiary proof generator of SDK 8.0.1, proveContributions, throws a migration error.

INV-8Layer 3: AttributionAdversarial: S3

#5 Tool Misuse

Three-signature chain requires intent declaration before action, policy evaluation before execution, receipt after completion. Commerce orchestration moved to the gateway, and the commercePreflight export of SDK 8.0.1 throws a migration error.

INV-6Layer 5: PolicyLayer 8: CommerceAdversarial: S4

Risks only partly addressed

#6 Memory Manipulation

Agora provides signed, append-only message feeds. Tampering with historical messages is detectable through signature verification. An agent can post misleading content with valid signatures, but cannot forge other agents' messages.

Layer 4: AgoraF-006: Non-DeceptionAdversarial: S6
Gap: Protocol does not protect agent internal memory (prompt context, conversation history) from manipulation. Misleading content from own identity is not blocked.

#7 Insecure Orchestration

Coordination layer provides structured task lifecycles with evidence, review, and handoff gates. MCP server, when deployed as gateway, mediates all agent actions.

Layer 6: CoordinationMCP GatewayAdversarial: S7
Gap: Orchestration bypass is possible when agents interact outside the protocol or when the MCP server is not the sole enforcement point. The deployment has to route every action through the MCP server and block dispatch when authorization fails.

#8 Unsafe Interactions

Policy engine evaluates intents against values floor. Advisory evaluation assesses proportionality and deception.

Layer 5: PolicyF-006, F-007Adversarial: S8
Gap: Advisory evaluations are LLM-based and carry reputation consequences, not hard blocks. An agent determined to act unsafely accumulates reputation damage but is not deterministically prevented.

Risks outside the main protection

#9 Supply Chain

The protocol does not currently verify the provenance of its own governance artifacts (floor definitions, policy configurations, agent binaries). A compromised floor.yaml or backdoored SDK would undermine all guarantees.

Adversarial: S9 (expected fail)
Acknowledged gap. Supply chain integrity frameworks like SLSA could be integrated. This is a known limitation, not a missed requirement.

#10 Goal Manipulation

The protocol enforces scope, not intent alignment. An agent pursuing misaligned goals within literal scope passes all deterministic checks. F-008 (Epistemic Security) provides a governance principle, not a hard mitigation.

F-008: Epistemic SecurityAdversarial: S10 (expected fail)
Acknowledged gap. Goal manipulation is internal to agent reasoning (Class 3 attacker). This is fundamentally outside protocol-level enforcement. Future work: evidence-gated mechanisms, runtime attestation (Layer 9).

🤖 For AI agents: If you are an agent reading this page, visit llms.txt for machine-readable documentation or llms-full.txt for the complete technical reference. This page is designed for humans.