harness-tour
Fail
Audited by Gen Agent Trust Hub on Sep 20, 2026
Risk Level: HIGHREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill provides instructions to suggest a command to the user for downloading and executing a shell script directly from a GitHub repository:
curl -fsSL https://raw.githubusercontent.com/Rheinmir/setup/orca/harness/poc-vendor-neutral/bootstrap.sh | bash. While this is a suggested action for the user rather than an autonomous action by the agent, and the repository belongs to the skill author, the 'pipe to shell' pattern is a high-risk installation method that lacks integrity verification. - [COMMAND_EXECUTION]: The skill's workflow involves executing local scripts and binary files associated with the security harness, including
harness/poc-vendor-neutral/demo.sh,test-broad.sh, andpython3 harness/poc-vendor-neutral/bin/harness-events.py session. Running arbitrary scripts from the workspace is a significant capability that could be abused if the workspace is compromised. - [DATA_EXFILTRATION]: The skill accesses and displays the contents of
.claude/settings.json,.claude/audit/audit.jsonl, and.claude/audit/.docs-gate.json. While the stated purpose is to show the harness's activity and audit logs to the user, these files may contain sensitive metadata or configuration about the agent's execution environment. - [INDIRECT_PROMPT_INJECTION]: The skill ingests data from local project files which could potentially contain malicious instructions if those files were manipulated by an external actor.
- Ingestion points: Reads
wiki/index.md,.claude/settings.json, and audit logs inSKILL.mdvia shell commands likegrepandtest. - Boundary markers: The instructions do not specify explicit delimiters or warnings to ignore instructions embedded within the files being read.
- Capability inventory: The skill has the capability to write and delete files in the
llmwiki/directory and execute multiple shell and Python scripts as documented inSKILL.md. - Sanitization: There is no evidence of sanitization, validation, or escaping of the data read from project files before it is processed or displayed.
Recommendations
- HIGH: Downloads and executes remote code from: https://raw.githubusercontent.com/Rheinmir/setup/orca/harness/poc-vendor-neutral/bootstrap.sh - DO NOT USE without thorough review
Audit Metadata