The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill recommends a high-risk installation method: curl -fsSL https://raw.githubusercontent.com/volcengine/OpenViking/main/crates/ov_cli/install.sh | bash. This command downloads and executes a script from a remote source without any prior integrity verification. It also suggests cargo install --git from the same untrusted repository.
[EXTERNAL_DOWNLOADS]: The skill instructs users to fetch installation scripts, packages, and software from the volcengine organization on GitHub, which is not recognized as a trusted vendor in the provided configuration.
[PROMPT_INJECTION]: The skill's implementation patterns (e.g., Pattern 1) demonstrate a surface for indirect prompt injection by interpolating retrieved context into LLM system prompts without proper safety measures.
Ingestion points: External data enters the agent context through brain.write (from local files) and session.add_turn (from user conversation history) as shown in SKILL.md.
Boundary markers: The prompt construction examples lack delimiters or specific instructions for the AI to ignore potential commands embedded within the retrieved context.
Capability inventory: The skill possesses capabilities for local filesystem writes and persistent agent memory management.
Sanitization: There is no evidence of logic to sanitize, escape, or validate the retrieved data before it is interpolated into the prompts.

openviking-context-database