oma-image
Fail
Audited by Gen Agent Trust Hub on Oct 6, 2026
Risk Level: HIGHCOMMAND_EXECUTIONPRIVILEGE_ESCALATIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill makes extensive use of subprocess calls to external command-line tools, including
oma image generate,codex exec, and theagy(Antigravity) CLI. These tools are central to the skill's operation but represent a broad execution surface. - [PRIVILEGE_ESCALATION]: The skill invokes the Antigravity CLI (
agy) with the--dangerously-skip-permissionsflag. This flag is explicitly designed to bypass security and permission checks within the tool's agentic loop, allowing it to perform actions without user confirmation. Additionally, thecodextool is called with--skip-git-repo-checkto suppress environment-level safety validations. - [EXTERNAL_DOWNLOADS]: The skill instructs the agent to fetch content from the
YouMind-OpenLabGitHub repositories (awesome-gpt-image-2andawesome-nano-banana-pro-prompts) using thegh api. These repositories are not within the trusted vendor scope and their content is not verified before being processed by the agent. - [INDIRECT_PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection through its prompt amplification process, which relies on untrusted remote data.
- Ingestion points: Remote README files from
YouMind-OpenLabare fetched viagh apiand decoded from Base64 into the agent's context. - Boundary markers: The instructions do not define boundary markers or safety delimiters for the remote content being ingested.
- Capability inventory: The skill possesses significant capabilities, including file system access and the ability to execute complex instructions via the
codexandagyagentic loops. - Sanitization: The agent is instructed to 'internalize the pattern' of the fetched content, but no sanitization, filtering, or validation is performed on the remote text to prevent the ingestion of malicious instructions.
Recommendations
- AI detected serious security threats
Audit Metadata