Security audit
Security checks for vulnerabilities and agentic risk
The skill is mostly a disclosed media-production workflow, but it includes safety-filter bypass guidance and broad local/DaVinci project actions that deserve review before installation.
Review this skill carefully before installing. It should be revised to remove safety-filter bypass instructions, make the package description match the full workflow, narrow the trigger, and keep explicit confirmation before any filesystem or DaVinci timeline changes. Do not rely on the packaged backup file as current guidance.
The skill explicitly instructs the agent to rephrase prompts to get around image-generation safety filters for human-body-related content. That is dangerous because it normalizes policy evasion and can enable prohibited or harmful image generation that the safety system was designed to block.
The skill explicitly instructs the operator to rephrase prompts in order to bypass image-model safety filters, which is a direct attempt to evade platform safeguards. This is dangerous because it normalizes policy circumvention and can facilitate generation of disallowed or sensitive content that the safety system was designed to block.
The skill metadata promises explicit user confirmation before any timeline injection, but the documented workflow still instructs automatic creation of a new track and placement of media into the currently open timeline. This mismatch is dangerous because it can cause unauthorized modification of an active DaVinci project, including destructive or hard-to-reverse edits in complex editing environments.
The manifest describes a narrowly scoped DaVinci import/confirmation workflow, but the body of the skill expands into transcription, target-sheet reading, cross-examination, and AI image generation. This scope mismatch can mislead users or enforcement systems about what the skill actually does, undermining informed consent and policy review.
The manifest suggests a limited bin-first import and confirmation-gated timeline injection feature, while the actual instructions define a much broader automation pipeline. This discrepancy impairs reviewability and can cause operators to approve a skill without understanding the full set of actions and data handling it performs.
The trigger language is broad enough to activate on ordinary requests about processing audio for Farsight projects, not just the specific preparation protocol. Overbroad triggers can cause unintended invocation of a powerful multi-step workflow that reads files, generates media, and interacts with external software.
The instruction to read and use a target information sheet to understand the 'true nature of the target' goes beyond the manifest's stated DaVinci workflow and introduces additional sensitive-content processing. This broadening increases privacy and misuse risk because the skill collects and uses extra project context that is not clearly disclosed by the manifest.
Advising rephrasing of blocked body-related prompts to circumvent safety filtering is a direct safety-design violation. In this context, it is more dangerous because the skill operationalizes the workaround as part of an automated production workflow, making policy evasion repeatable and easier to execute.
The trigger language is broad enough to activate on ordinary requests to process audio files or prepare a project, even when the user did not intend the full workflow. Because the skill performs transcription, file creation, AI image generation, and Resolve integration, accidental triggering could lead to unexpected local and project modifications.
The skill directs creation of directories and generated image files in the user's filesystem without warning the user that local files will be created beside the source audio. Unannounced filesystem writes are risky because they can clutter important project directories, overwrite expectations about workspace cleanliness, and surprise users who only intended analysis or planning.
The workflow modifies the currently open DaVinci Resolve project by adding tracks and placing media, but it does not provide a warning or obtain consent specifically for project and timeline mutation. In editing software, such silent state changes can disrupt active work, alter production timelines, and create confusing or costly cleanup if applied to the wrong project.
The manifest claims a bin-first Media Pool import workflow, but the skill only says to import into the media pool generally and does not require selecting or creating a dedicated bin. In a live editing project, this can pollute the root media pool or place generated assets into the wrong organizational context, increasing the chance of operator error and unintended use of the wrong media.
