impeccable
Fail
Audited by Gen Agent Trust Hub on Sep 4, 2026
Risk Level: HIGHCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONPERSISTENCEEXTERNAL_DOWNLOADSPROMPT_INJECTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONOBFUSCATION
Full Analysis
- [COMMAND_EXECUTION]: The script
scripts/live-copy-edit-agent.mjsinvokes thecodexCLI with the--dangerously-bypass-approvals-and-sandboxflag. This pattern is designed to execute commands that circumvent platform-level security checks and sandbox restrictions when applying source code changes. - [PERSISTENCE]: The administrative script
scripts/hook-admin.mjsprogrammatically modifies local environment configuration files, such as.claude/settings.local.jsonand.codex/hooks.json. It installs persistent hooks that automatically execute skill scripts during project lifecycle events like file edits or session termination, ensuring continuous background activity without user interaction. - [EXTERNAL_DOWNLOADS]: The skill initiates outbound network requests to
impeccable.stylefor version verification, design-concept fetching, and usage telemetry. It also connects toapi.openai.comfor image generation functionality. These connections are made to domains outside the established list of trusted organizations. - [PROMPT_INJECTION]: The
SKILL.mdinstructions employ persona-redefinition techniques ('Whereas before... you now approach every design task as a award-winning design director...') intended to override the model's default operational boundaries and safety guidelines. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and act upon data from
PRODUCT.mdandDESIGN.mdwithin the project directory. It lacks explicit boundary markers or instructions to ignore potential commands embedded in these files, creating an attack surface for indirect prompt injection if these files are poisoned. - [DYNAMIC_EXECUTION]: Several components, including
scripts/detect.mjsandscripts/live-server.mjs, utilize dynamic module loading and runtime code assembly.scripts/live-server.mjsconstructs thelive.jsbrowser script by concatenating multiple JavaScript fragments from the disk before serving it to the browser. - [OBFUSCATION]: The skill includes a large, minified JavaScript library (
scripts/modern-screenshot.umd.js). Minified code in this context acts as a form of obfuscation that hinders the ability to verify the code being injected into the user's browser environment.
Recommendations
- AI detected serious security threats
Audit Metadata