concept-guard
Pass
Audited by Gen Agent Trust Hub on Nov 1, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_CONTEXT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill implements a mechanism to automatically read and inject architectural constraints from local markdown files into the agent's reasoning context during file edits.
- Ingestion points: The
runtime/scripts/drift-context.shandruntime/scripts/post-check.shscripts read sections fromCONCEPT.mdandPIPELINE.mdfiles located in the project directory when the agent initiates a write or edit operation. - Boundary markers: Injected context is prefixed with the file path and section name (e.g.,
[relative/path/to/spec ## interactions]) but lacks instructions to the model to disregard potential instructions embedded within those data sections. - Capability inventory: The agent has permissions to read/write/search files and execute shell commands within its environment.
- Sanitization: The scripts extract raw text from markdown sections using
sedwithout filtering for potential prompt injection patterns before passing it to the agent's prompt context. - [DYNAMIC_CONTEXT_INJECTION]: The skill utilizes the Claude Code hook system to execute bash scripts that dynamically provide context to the agent.
- Evidence:
runtime/hooks/hooks.jsondefines hooks forSessionStart,PreToolUse, andPostToolUsethat execute the provided bash scripts. - Logic:
drift-context.shperforms an upward directory traversal to find architectural specifications relevant to the file being edited, injecting their content directly into the tool call's output asadditionalContext. - [COMMAND_EXECUTION]: The skill relies on local shell script execution for its core monitoring and context-injection features.
- Evidence: The scripts
runtime/scripts/session-start.sh,runtime/scripts/drift-context.sh, andruntime/scripts/post-check.shuse standard Unix utilities likefind,jq,sed, andawkto analyze the project state and report it to the agent.
Audit Metadata