run-feature
Fail
Audited by Gen Agent Trust Hub on Nov 8, 2026
Risk Level: HIGHPRIVILEGE_ESCALATIONCOMMAND_EXECUTIONCREDENTIALS_UNSAFEEXTERNAL_DOWNLOADSPROMPT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PRIVILEGE_ESCALATION]: The skill explicitly bypasses agent security filters by using the
--dangerously-skip-permissionsflag for theclaudeCLI and requestingfull-accesspermission modes for sessions launched via the local daemon. - [COMMAND_EXECUTION]: The scripts (
run.mjsanddaemon-session.mjs) execute a wide range of shell commands, includinggit,gh,osascript, and various package managers, with high autonomy and no human-in-the-loop verification for individual operations. - [CREDENTIALS_UNSAFE]: The skill accesses sensitive authentication material, specifically GitHub App private keys located at
~/.config/agents-hq/github-app.pemand App IDs to mint temporary tokens for repository access. - [EXTERNAL_DOWNLOADS]: The runner script automatically triggers package installation commands (such as
pnpm installandnpm ci) in generated worktrees when it detects corresponding lockfiles, leading to the download and execution of code from external registries without prior verification. - [PROMPT_INJECTION]: The skill injects instructions into sub-sessions (
HEADLESS_NOTE) that direct the agent to skip human confirmation and ignore questions, prioritizing autonomous task completion over interactive safety protocols. - [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted project data (e.g.,
decisions.md,spec.md,slices.json) and uses it to drive automated sessions with high privileges. This creates a surface where malicious instructions in project files could be executed with full host access.
Recommendations
- AI detected serious security threats
Audit Metadata