The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted token metadata, such as names, symbols, and descriptions, from external providers. This data is utilized to parameterize sensitive financial operations, including token transfers and swaps, which creates a potential surface for indirect prompt injection.
Ingestion points: Data from external sources is introduced into the agent context via the onchain_search_tokens, onchain_get_token_metadata, and onchain_get_address_portfolio tools.
Boundary markers: The instructions do not define explicit delimiters to isolate external content, but they do mandate a human-in-the-loop confirmation process for all transaction attempts.
Capability inventory: The agent is empowered to execute blockchain transactions through the onchain_send_token and onchain_jupiter_execute tools.
Sanitization: The skill incorporates mitigation strategies including required confirmation for transfers, specific verbatim warnings for external addresses, and the mandatory display of market data and verification links (DexScreener) for project assessment.

openfin-onchain