The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes a framework where agents discover and interact with other entities based on external, potentially attacker-controlled data.
Ingestion points: The agent is instructed to fetch and parse agentURI JSON files from IPFS or web endpoints and read ReputationRegistry data from blockchain contracts.
Boundary markers: No boundary markers or instructions to ignore embedded prompts are present in the integration examples.
Capability inventory: The skill utilizes x402Fetch for network operations and ethers.Contract for executing on-chain transactions.
Sanitization: No sanitization or validation of the external metadata is described, creating a risk where malicious registration files could influence agent logic.
[EXTERNAL_DOWNLOADS]: The skill references several external packages for implementing agent commerce.
Evidence: Mentions Node.js packages @x402/express, @x402/fetch, @x402/evm, @x402/core, and ethers, as well as a Python package x402. These are associated with well-known Ethereum development tools and services.
[METADATA_POISONING]: The skill uses deceptive future dates to establish authority for protocols that are not part of current standard training data.
Evidence: The document claims ERC-8004 was deployed in January 2026 and EIP-7702 was shipped in May 2025. This hallucinated authority may attempt to override the model's existing knowledge of Ethereum standards.

standards