Decagon — agentic threat model

Not certain from the listing — Likely relies on proprietary or fine-tuned commercial foundation models. Primary threats include prompt injection attacks that bypass customer service guidelines, leading to brand damage or unauthorized instructions.

Not certain from the listing — Integrates with enterprise data sources for RAG. Threats include knowledge-base poisoning where malicious data injected into connected systems alters the agent's behavior, and accidental exfiltration of customer PII.

Utilizes customizable workflows and seamless integrations to orchestrate actions. Threats include tool misuse where the agent is manipulated into executing unauthorized API calls (e.g., processing fraudulent refunds or modifying account settings).

Not certain from the listing — Deployed as an enterprise SaaS platform. Threats include insecure API endpoints, lack of tenant isolation, and exposure of integration secrets/credentials.

Features 'analytics and insights' and 'transparent decision-making' to monitor agent behavior. However, threats remain regarding blind spots in detecting sophisticated, multi-turn adversarial prompt injections or drift in continuous learning loops.

Not certain from the listing — While targeting enterprise customers, specific compliance certifications (e.g., SOC2, GDPR) are not detailed. Threats include insufficient authorization checks between the agent and backend enterprise systems.

Not certain from the listing — Primarily functions as a direct customer-to-agent interface, but integration with external tools creates ecosystem dependencies. Threats include cascading failures if third-party APIs or integrated services are compromised.