AgentRunner — agentic threat model

Not certain from the listing — the platform supports various external AI models, making it susceptible to model-specific threats like adversarial prompt injection, model stealing, or misaligned outputs depending on which third-party LLMs developers integrate.

Not certain from the listing — data operations are not detailed, but as an orchestration platform, it likely handles sensitive training, testing, or RAG data, risking data exfiltration or poisoning if integrations are insecure.

As an orchestration platform, L3 is highly critical. Vulnerabilities include insecure tool integration, tool misuse within visual workflows, and framework-level flaws that could allow malicious agents to execute unauthorized actions.

Not certain from the listing — while it claims 'scalable deployment infrastructure', specific sandboxing, container isolation, or secrets management practices are not detailed, posing risks of container escape or privilege escalation.

Strongly supported. The platform features dedicated testing, evaluation, and real-time monitoring frameworks, which help mitigate blind spots and drift, though they must be secured against evaluation gaming.

Not certain from the listing — no specific security certifications, RBAC, or compliance frameworks (like SOC2 or GDPR) are mentioned, leaving identity and authorization controls unverified.

The platform orchestrates multiple agents and workflows, introducing risks of cascading failures, agent-to-agent trust abuse, and rogue agent behavior within complex visual pipelines.