Storyloft — agentic threat model

Not certain from the listing — likely relies on commercial third-party LLMs to power the 'Eddy' assistant. Primary threats include prompt injection to bypass content filters or extract system prompts, and potential data leakage if the underlying model provider uses manuscript inputs for training.

Not certain from the listing — requires storing large manuscripts, chapters, and notes, likely utilizing a cloud database or vector store to enable 'manuscript-aware' context. Threats include unauthorized access to user intellectual property, data exfiltration, and potential database poisoning if malicious inputs are saved into the manuscript state.

Not certain from the listing — Eddy orchestrates manuscript reading, editing, and formatting. Threats include insecure handling of large context windows (e.g., indirect prompt injection via manuscript content) and potential tool misuse if formatting or illustration generation tools are exposed to direct LLM manipulation.

Not certain from the listing — hosted as a closed-source SaaS platform. Threats include standard web application vulnerabilities, insecure API endpoints, and lack of strict tenant isolation which could allow cross-user manuscript access.

Not certain from the listing — no mention of monitoring, guardrails, or evaluation frameworks for Eddy's outputs. Gaps could lead to undetected generation of offensive content or silent corruption of manuscript files during automated editing.

Not certain from the listing — as a paid, closed-source vertical app, it must protect user IP, but no specific compliance certifications (e.g., SOC 2, GDPR) or robust access controls are detailed in the public directory.

Not certain from the listing — operates as a standalone vertical writing application with no apparent multi-agent marketplace or external agent-to-agent integrations, minimizing ecosystem-level threats.