The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill directs the download of instructions and guidance files from an unverified GitHub repository (Copenhagen0x/solana-security-standard) into sensitive directories such as .claude/ and .cursor/. Because these files are intended to control the AI agent's behavior, a compromise of the remote repository could result in the agent being hijacked by malicious instructions.
[EXTERNAL_DOWNLOADS]: The skill makes extensive use of npx to execute code directly from the NPM registry (@jelleo/solana-security-standard and @jelleo/solana-security-mcp). These executions occur without version pinning or verification of the third-party publisher's identity.
[PERSISTENCE]: The skill encourages the creation of a persistent execution mechanism by writing a script to .git/hooks/pre-commit. This script is designed to run the unverified @jelleo security scanner automatically every time a user attempts to commit code to their repository.
[INDIRECT_PROMPT_INJECTION]: The skill defines tools and workflows for scanning arbitrary Solana program code, creating a surface for indirect prompt injection attacks.
Ingestion points: The scan_solana_code MCP tool and CLI scan commands ingest external source code for analysis as described in SKILL.md.
Boundary markers: There are no explicit instructions or delimiters provided to ensure the agent ignores natural language instructions that might be embedded within the code files being scanned.
Capability inventory: The skill environment permits filesystem writes (to agent config and git hooks), network requests (via curl), and arbitrary command execution (via npx).
Sanitization: The skill does not document any sanitization or validation logic to prevent malicious code snippets from influencing the agent's logic during the scanning process.

solana-security-standard