The Agent Skills Directory

[DATA_EXFILTRATION]: The script scripts/audit_hindsight_memory.py attempts to access a configuration file at the hardcoded absolute path /home/delorenj/.openclaw/openclaw.json. This pattern exposes sensitive user-specific configuration and reflects poor security practices for a distributed skill.
[REMOTE_CODE_EXECUTION]: Documentation and setup guides repeatedly recommend executing remote shell scripts directly via piped commands (e.g., curl ... | bash). Specifically, scripts for the CLI, MCP server, and agent skill are fetched from hindsight.vectorize.io and executed without verification.
[DYNAMIC_EXECUTION]: The skill configuration supports the HINDSIGHT_API_EMBEDDINGS_LOCAL_TRUST_REMOTE_CODE environment variable, which enables the loading and execution of untrusted Python code embedded within machine learning models downloaded from the internet.
[INDIRECT_PROMPT_INJECTION]: The skill's architecture is inherently vulnerable to indirect prompt injection.
Ingestion points: Arbitrary content is ingested via hindsight memory retain as shown in SKILL.md and references/developer/api/retain.md.
Boundary markers: Explicit delimiters for ingested instructions are absent.
Capability inventory: The skill uses the hindsight CLI which has network and file system access capabilities.
Sanitization: External content is not sanitized before interpolation into prompts during recall or reflect operations.

hindsight