craft-handoff

Warn

Audited by Gen Agent Trust Hub on Sep 30, 2026

Risk Level: MEDIUMPERSISTENCEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PERSISTENCE]: The skill provides instructions and a script (load-pending-hook.mjs) to be installed as a SessionStart hook in ~/.claude/settings.json. This mechanism ensures that the script runs automatically every time a new session is started or context is cleared, effectively establishing a persistent execution point within the agent's environment.
  • [INDIRECT_PROMPT_INJECTION]: The auto-load mechanism creates a vulnerability surface for indirect prompt injection. The hook script reads and injects content from files located in ~/.craftkit/handoff/pending/ directly into the agent's context. A malicious local process could place a crafted file in this directory to influence the agent's behavior in future sessions.
  • Ingestion points: load-pending-hook.mjs reads content from files in ~/.craftkit/handoff/pending/.
  • Boundary markers: The prompt format uses XML-like tags (<context>, <task>, <rules>) to delimit sections, but these do not prevent an adversary from including instructions within those blocks.
  • Capability inventory: The skill allows the agent to execute git commands, write files to the home directory, and interact with the system clipboard.
  • Sanitization: While the skill instructions tell the agent to redact secrets before writing artifacts, the automated hook script does not perform sanitization or validation of the content it injects.
  • [COMMAND_EXECUTION]: The skill utilizes several bundled scripts to perform its tasks. gather-state.mjs executes git commands via execFileSync to collect repository metadata. copy-clipboard.sh is a shell script that attempts to execute various system clipboard utilities (pbcopy, wl-copy, xclip, xsel, clip.exe) to store the handoff prompt.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Sep 30, 2026, 12:27 AM