bootstrap
Pass
Audited by Gen Agent Trust Hub on Nov 7, 2026
Risk Level: SAFECREDENTIALS_UNSAFEDATA_EXFILTRATIONMETADATA_POISONINGINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [CREDENTIALS_UNSAFE]: The skill attempts to retrieve GitHub authentication tokens from the environment or by executing the
gh auth tokencommand to authenticate repository operations. - Evidence:
GH_TOKEN="${GH_TOKEN:-${GITHUB_TOKEN:-$(gh auth token 2>/dev/null || echo "")}}}"inSKILL.md. - [DATA_EXPOSURE]: The skill interacts with sensitive file paths by attempting to remove SSH control sockets to manage connection multiplexing.
- Evidence:
rm -f ~/.ssh/control-*andssh -O exit git@github.cominSKILL.md. - [DYNAMIC_EXECUTION]: The skill generates and writes a shell script (
tmp/bootstrap-claude-session.sh) to the local file system and changes its permissions to executable. - Evidence:
Phase 4: Generate Bootstrap ScriptinSKILL.mdusingcat > "$OUTPUT_FILE"andchmod +x "$OUTPUT_FILE". - [METADATA_POISONING]: The skill contains "Self-Evolving Skill" instructions that direct the AI to modify the
SKILL.mdfile itself to address issues or drifted parameters. - Evidence: "Self-Evolving Skill: This skill improves through use. If instructions are wrong... fix this file immediately" in
SKILL.md. - [INDIRECT_PROMPT_INJECTION]: The skill processes repository URLs and branch names from the local environment and user arguments, which are then interpolated into shell scripts used for cloning and pushing data.
- Ingestion points:
git remote get-url originand user-supplied arguments (-r,-b). - Boundary markers: None identified in the prompt interpolation.
- Capability inventory: Shell execution (
Bash), file writing (Write), and network operations (git push). - Sanitization: Uses standard shell variable expansion without explicit escaping for potential command injection in repository names.
Audit Metadata