Fluidworks — agentic threat model

Not certain from the listing — likely utilizes proprietary or third-party speech-to-text, text-to-speech, and LLM models. The primary threat is voice-based or text-based prompt injection that could hijack the model's output to trigger unintended UI actions.

Not certain from the listing — trained on the host application's UI, use cases, and customer questions, and continuously learns from real sessions. Threats include data exfiltration of sensitive user session data and poisoning of the feedback loop used for auto-learning.

The agent framework translates natural language intent directly into UI actions (clicking, navigating, filling fields). The critical threat is insecure tool execution, where the UI automation engine executes malicious actions (e.g., submitting forms or deleting data) due to adversarial instructions.

Not certain from the listing — likely deployed as an embedded JavaScript SDK or iframe within the host application. Threats include DOM-based vulnerabilities, cross-site scripting (XSS) via the agent's interface, and interception of real-time voice streams.

Not certain from the listing — features 'auto-learning and insights' to identify user drop-offs and FAQs, but does not specify security-focused guardrails, input filtering, or anomaly detection to block malicious UI manipulation attempts.

Not certain from the listing — no security certifications (e.g., SOC2, ISO 27001) or compliance frameworks are mentioned. There is a risk of the agent bypassing application-level authorization checks if it inherits user sessions without strict boundary controls.

Not certain from the listing — operates as a standalone onboarding assistant within a single application. No multi-agent coordination or external marketplace integrations are described.