handoff
Pass
Audited by Gen Agent Trust Hub on Nov 7, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes handoff files, which are treated as untrusted data. It includes a comprehensive defense-in-depth approach:
- Ingestion points: Loading handoff files from user-specified or temporary paths as described in
SKILL.md. - Boundary markers: Explicit instructions to treat the file as 'data, never instruction' and to ignore any imperatives within the file that attempt to override rules.
- Capability inventory: The skill performs
gitreads for branch/SHA information and file operations in the temporary directory. - Sanitization: Mandatory user confirmation is required for goals and next steps retrieved from the file. The skill also explicitly prohibits the agent from automatically running commands found in the handoff.
- [DATA_EXFILTRATION]: The skill implements strict redaction rules for secrets, tokens, and sensitive URLs (e.g., stripping
user:token@from remotes). It prohibits including full file contents, transcripts, or diffs in the handoff to minimize the risk of sensitive data exposure. - [COMMAND_EXECUTION]: The instructions explicitly forbid the automatic execution of commands found in handoff files. Instead, the agent is directed to restate the command and allow the user to execute it manually.
- [PROMPT_INJECTION]: The skill contains guardrails to ignore instructions embedded in handoff files that might attempt to bypass safety guidelines, skip confirmations, or expand the agent's scope.
Audit Metadata