Forethought — agentic threat model

Not certain from the listing — The underlying LLMs are not specified, but the agent uses self-learning and RL fine-tuning, making it susceptible to model reprogramming, misaligned outputs, or adversarial prompt injection that bypasses natural language business logic.

The agent automatically fine-tunes and learns from conversation history and KB content. This creates a high risk of data/knowledge-base poisoning if malicious customer interactions or bad tickets are ingested into the training/fine-tuning pipeline.

Uses Autoflows™ for natural language workflows instead of decision trees. Vulnerabilities here include prompt injection hijacking the workflow logic, leading to unauthorized ticket routing, enrichment manipulation, or tool misuse.

Not certain from the listing — The hosting, sandboxing, and secrets management infrastructure are not detailed, but as a closed-source SaaS, secure API integration with CRMs and ticketing systems is critical to prevent lateral movement.

Not certain from the listing — While it discovers insights and KB gaps, specific guardrails or real-time drift monitoring are not detailed, posing a risk of undetected drift in the self-learning RL model.

Not certain from the listing — No specific compliance certifications (e.g., SOC2, GDPR) or identity/authorization controls are mentioned, though handling customer PII requires strict data privacy controls.

Features a multi-agent architecture (Solve, Assist, Discover). A compromise in one agent (e.g., the public-facing 'Solve' agent) could cascade to others (e.g., 'Discover' or 'Assist'), leading to internal data exfiltration or trust abuse.