tao-generate-referring-expressions
Pass
Audited by Gen Agent Trust Hub on Oct 29, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill operates by ingesting and processing untrusted external data (images and KITTI-format labels) through a Vision-Language Model (VLM), creating a potential surface for instructions embedded in the data to influence the agent.
- Ingestion points: Untrusted data enters the agent context from
data.image_dir(images),data.kitti_label_dir(KITTI labels), anddata.input_annotations_jsonl(pre-existing annotation files). - Boundary markers: The documentation does not specify the use of clear delimiters or system instructions designed to prevent the VLM from following commands potentially hidden in the visual data or label text.
- Capability inventory: The skill utilizes shell command execution via the
auto_labeltool and has write permissions to the local filesystem for generating results. - Sanitization: There is no evidence of pre-processing, validation, or sanitization of the input text or image content before it is interpolated into VLM prompts.
Audit Metadata