SciFig — agentic threat model

Not certain from the listing — likely utilizes multimodal foundation models (e.g., vision-language models or custom diffusion architectures) to interpret sketches, PDFs, and text. Threats include adversarial inputs (poisoned sketches/PDFs triggering misaligned or offensive outputs) and model reprogramming.

Not certain from the listing — requires ingestion of user-uploaded PDFs, photos, and sketches. Threats include data exfiltration of proprietary/unpublished research data, and potential training data poisoning if user uploads are used for continuous fine-tuning without sanitization.

Not certain from the listing — likely uses a lightweight orchestration framework to parse inputs, generate design suggestions, and render editable vector/raster outputs. Threats include insecure file parsing (PDF/image exploits) and prompt injection via input text or PDF metadata.

Not certain from the listing — likely hosted as a web application (given 'Paid' and 'Open Source' tags). Threats include server-side request forgery (SSRF) if it fetches external reference images, and container escape/resource exhaustion during heavy image rendering tasks.

Not certain from the listing — lacks explicit mention of guardrails or output verification. Threats include a lack of automated validation for scientific accuracy in generated figures, leading to undetected hallucinations or distorted data representation.

Not certain from the listing — no mention of compliance standards (e.g., GDPR, SOC2) or access controls for sensitive, unpublished academic research. Threats include unauthorized access to user-generated figures and intellectual property theft.

Not certain from the listing — operates primarily as a standalone horizontal tool without explicit multi-agent or marketplace integrations. Threats are minimal here, but could emerge if integrated into broader academic publishing workflows.