The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides numerous examples and instructions for fetching, monitoring, and decoding data from the Solana blockchain (e.g., account data, transaction logs, and WebSocket notifications). Since blockchain data is provided by external, untrusted actors, it represents a surface for indirect prompt injection attacks if the agent processes this data without adequate sanitization.
Ingestion points: Examples in examples/subscriptions/README.md (account and log notifications) and examples/fetch-accounts/README.md (encoded account fetching) demonstrate reading external blockchain state into the agent's context.
Boundary markers: The provided code examples do not include explicit instructions, delimiters, or markers to distinguish between trusted instructions and untrusted data fetched from the chain, increasing the risk of the agent following instructions embedded in transaction memos or account metadata.
Capability inventory: The skill enables the agent to perform network requests (RPC calls), read local files (e.g., loading keys via fs.readFileSync in examples/transfer-sol/README.md), and potentially execute shell commands if the agent chooses to run the provided scripts for transaction submission.
Sanitization: The documentation focuses on functional usage of the SDK and does not implement or mandate data sanitization for the fetched blockchain content before it is processed by the LLM.

solana-kit