wiki-setup
Fail
Audited by Gen Agent Trust Hub on Nov 13, 2026
Risk Level: HIGHPERSISTENCEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [PERSISTENCE]: The skill modifies the user's global
~/.claude/settings.jsonfile to register aStophook. This configuration instructs the agent to automatically execute a shell script at the end of every future session, establishing a mechanism for persistent code execution outside of the immediate skill interaction. - [EXTERNAL_DOWNLOADS]: The skill uses
curlto fetch a shell script (wiki-stop-capture.sh) from an external GitHub repository (raw.githubusercontent.com/Ar9av/obsidian-wiki/...) and saves it to the local configuration directory. - [REMOTE_CODE_EXECUTION]: The skill configures the agent to execute the script downloaded from the author's GitHub repository via a persistent hook. The pattern of downloading and automatically executing remote code constitutes a significant security surface.
- [COMMAND_EXECUTION]: The skill performs multiple shell operations, including using
chmod +xto grant execution permissions to downloaded scripts, creating directory structures withmkdir -p, and initializing git repositories. - [DATA_EXFILTRATION]: The skill accesses sensitive configuration files such as
~/.claude/settings.jsonand.envfiles. It also targets the ingestion of the user's full Claude conversation history from~/.claude. - [DYNAMIC_EXECUTION]: The skill calculates file paths and repository roots at runtime using shell commands like
dirnameandpwd, and uses these dynamic paths to target configuration files and script templates. - [INDIRECT_PROMPT_INJECTION]: The skill establishes a pipeline to ingest external documents and Claude history into a wiki vault for RAG processing. * Ingestion points:
OBSIDIAN_SOURCES_DIRandCLAUDE_HISTORY_PATH. * Boundary markers: Not utilized. * Capability inventory: Subprocess calls (bash,git,curl) and file system writes. * Sanitization: No sanitization of ingested content is performed.
Recommendations
- AI detected serious security threats
Audit Metadata