hermes-lcm-context-management

Fail

Audited by Gen Agent Trust Hub on May 17, 2026

Risk Level: HIGHEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATIONREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill documentation instructs users to clone a repository from an unverified third-party account (https://github.com/stephenschoettler/hermes-lcm) to the local environment.\n- [COMMAND_EXECUTION]: The skill requires the execution of shell scripts (scripts/install.sh, scripts/update.sh) provided within the unverified external repository, which could lead to arbitrary code execution.\n- [DATA_EXFILTRATION]: The skill aggregates conversation history, which may contain sensitive user data or secrets, into a local database and subsequently transmits this data to external AI models (Anthropic Claude and OpenAI GPT) for summarization and expansion.\n- [REMOTE_CODE_EXECUTION]: Combining unverified repository cloning with the execution of provided scripts enables a remote code execution vector if the source repository is compromised.\n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it reads and processes external chat history and summaries via tools like lcm_grep and lcm_expand_query.\n
  • Ingestion points: Raw messages and summaries stored in ~/.hermes/profiles/<profile>/lcm.db (SKILL.md).\n
  • Boundary markers: None mentioned for identifying or ignoring instructions within retrieved history.\n
  • Capability inventory: Uses lcm_expand_query to call auxiliary LLMs and potentially other shell commands via slash commands (SKILL.md).\n
  • Sanitization: No evidence of content escaping or validation of retrieved historical data before re-interpolation into prompts.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
May 17, 2026, 08:19 PM
Security Audit — agent-trust-hub — hermes-lcm-context-management