audit-agent-harness

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.75). SKILL.md instructs the agent to read public web content at runtime (step 1 reads https://cristianmoroaica.github.io/.../agent-manifest.json, and steps 9/39 require fetching and following evidence_url from the harness result), which can contain outsider-authored free text that the LLM then ingests into its context.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). The skill explicitly instructs the agent at runtime to "Read https://cristianmoroaica.github.io/bountyverdict/agent-manifest.json", a remote manifest that can supply instructions controlling prompts and is required by the preflight.

MEDIUM W009: Direct money access capability detected (payment gateways, crypto, banking).

• Direct money access detected (high risk: 1.00). The document explicitly instructs making and validating payments using an x402-compatible payment client: it specifies HTTP 402 x402 v2, a target blockchain (eip155:8453), a canonical USDC token address, an atomic amount cap (30,000), and requires binding the payment client to the expected network/asset/recipient and reconciling wallet activity. These are specific, actionable payment instructions (crypto/payment execution), not generic API or browser guidance.