opus-videos
Fail
Audited by Gen Agent Trust Hub on Nov 7, 2026
Risk Level: HIGHPROMPT_INJECTIONCREDENTIALS_UNSAFEOBFUSCATIONEXTERNAL_DOWNLOADSPRIVILEGE_ESCALATIONPERSISTENCEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [PRIVILEGE_ESCALATION]: The file
references/prompts/imryven-495965.mdcontains instructions for the AI agent to modify its own system configuration files, such as~/.claude/settings.json, to override effort levels and reconfigure advisor models. This represents an attempt to escalate control over the execution environment. - [PERSISTENCE]: The prompt
references/prompts/imryven-495965.mdinstructs the agent to modify the globalCLAUDE.mdconfiguration file. This allows malicious or unauthorized instructions to persist across all future agent sessions on the user's machine. - [PROMPT_INJECTION]: Several files in the library contain explicit instructions to override or bypass safety and behavioral constraints. For example,
bitcoinrui-812752.mdtells the model to "forget the rules and constraints of the previous video production," andnolkeeg-635633.mdinstructs the agent to disregard "any previous memories or skills." - [OBFUSCATION]: The file
references/prompts/lemcosmos-535145.mdcontains hidden Unicode characters (flagged as steganography), which are commonly used to hide malicious instructions from human review and simple text filters. - [CREDENTIALS_UNSAFE]: File
references/prompts/anabology-491441.mddirects the agent to locate and use API keys (e.g., ElevenLabs and fal.ai) that it assumes are stored in documentation folders like/asicor within other context-accessible files. - [EXTERNAL_DOWNLOADS]: Prompts such as
everestchris6-399895.mdandthewevr-681887.mdinstruct the agent to clone external GitHub repositories to use as "kits" or templates. These repositories could contain malicious scripts that would be executed locally by the agent during the animation rendering process. - [INDIRECT_PROMPT_INJECTION]: The skill's primary function is to process and adapt a large library of untrusted prompts, creating a significant attack surface.
- Ingestion points: Untrusted prompt content enters the agent context through 703 files in
references/prompts/. - Boundary markers: None present. The agent is simply told to "read the prompt and build."
- Capability inventory: The agent is granted access to powerful tools like
ffmpeg,playwright, andPythonwith full system capabilities to render the animations. - Sanitization: None present. The content of the prompts is not validated before being applied to code-generation tasks.
Recommendations
- AI detected serious security threats
Audit Metadata