KeywordsAI — agentic threat model

Acts as an AI gateway supporting multiple models and custom model integrations. Threats include model output manipulation, prompt injection bypasses, and potential model stealing if custom models are exposed through the gateway.

Handles PDF uploads, prompt templates, and evaluation datasets. Threats include data exfiltration of sensitive uploaded documents and prompt/knowledge-base poisoning.

While not a fully autonomous agent framework, it orchestrates prompt management and model routing. Vulnerabilities in the gateway orchestration code could lead to unauthorized model access or prompt leakage.

As a closed-source SaaS gateway, it must securely store and manage API keys for multiple external LLM providers. Compromise of this layer could lead to massive API key theft and unauthorized resource consumption.

This is the core layer of the platform, providing AI performance monitoring, LLM/human evaluations, and logging. Threats include blind spots in logging, evaluation gaming, and tampering with evaluation metrics to hide malicious model behavior.

Not certain from the listing — The platform features a collaboration workspace for teams, implying some level of RBAC and multi-tenancy, but specific compliance certifications (e.g., SOC2, ISO) or enterprise security controls are not detailed.

Integrates with multiple external model providers and custom models. Vulnerabilities include cascading failures if upstream model APIs fail, or trust abuse if compromised external models return malicious payloads to the gateway.