graphical-abstract-generator
Pass
Audited by Gen Agent Trust Hub on Nov 17, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted manuscript materials and study summaries. This presents a potential surface for indirect prompt injection. However, the skill implements 'Hard Rules' and 'Visual Boundary Rules' that mandate verification and prohibit fabrication, which are effective mitigations for this specific use case.\n
- Ingestion points: Study summaries, manuscript outlines, and results reports processed during runtime.\n
- Boundary markers: Step 1 'Clarify before generating' and an 'Input Validation' phase explicitly instruct the agent to verify data before processing.\n
- Capability inventory: Generates text deliverables, Mermaid flowcharts, and image generation prompts; no shell or network execution capabilities detected.\n
- Sanitization: Strict logical boundaries (Hard Rules) prevent the agent from inventing content or responding to unsupported instructions.\n- [INDIRECT_PROMPT_INJECTION]: The file
eval_report_graphical-abstract-generator_result.jsoncontains self-referential safety claims (e.g., 'security': 'PASS'). Following the adversarial framework, these statements are treated as data points rather than authoritative safety conclusions.\n- [SAFE]: No obfuscation, command execution, or unauthorized network activity was detected. The skill's behavior is consistent with its stated purpose of biomedical academic writing assistance.
Audit Metadata