llm-wiki
Pass
Audited by Gen Agent Trust Hub on Nov 10, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATIONCOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and distill external data sources (documents, research papers, web articles, and images via vision-capable models) into a persistent knowledge base. This creates an attack surface where instructions embedded in processed sources could influence agent behavior during ingestion or subsequent query operations.\n
- Ingestion points: Defined in
SKILL.md(Raw Sources Layer 1) and includes local source directories, scratch inbox folders, and external URLs.\n - Boundary markers: The instructions do not specify the use of delimiters or 'ignore embedded instructions' warnings when interpolating source content into prompts.\n
- Capability inventory: The skill relies on shell-based interaction via custom CLI tools (
obsidian-wiki,qmd) and standard utilities (rg,grep,find) to manage the file system.\n - Sanitization: No specific filtering or sanitization logic is described for the raw content being processed.\n- [DATA_EXFILTRATION]: The skill's configuration protocol instructs the agent to recursively search for
.envfiles in all parent directories up to the user's home directory ($HOME). This pattern exposes high-value environment files, which often contain secrets, to the agent's context during the search for vault configuration.\n - Evidence: The
find_configfunction logic inSKILL.mdexplicitly directs the agent to walk up the directory tree to$HOMEand read.envfiles to find theOBSIDIAN_VAULT_PATHvariable.\n- [COMMAND_EXECUTION]: The skill requires the use of several shell utilities for searching, indexing, and maintaining the integrity of the knowledge vault.\n - Evidence: Usage of
ripgrep(rg),grep,find,md5sum, andmd5is documented acrossSKILL.mdfor retrieval and state-scoping tasks.\n- [DYNAMIC_EXECUTION]: The skill provides functional logic (bash-style) for configuration resolution and state management that the agent is expected to execute or implement during its lifecycle.\n - Evidence: The 'Config Resolution Protocol' in
SKILL.mdand the transaction/lock management logic inreferences/MEMORY.md.
Audit Metadata