Neo — agentic threat model

Not certain from the listing — the underlying foundation models used by Neo are not specified. However, threats include adversarial prompt injection bypassing safety guardrails to execute malicious code during feature engineering or model deployment.

Neo ingests raw data for preprocessing, training, and evaluation. Threats include training data poisoning (leading to backdoored models) and data exfiltration of sensitive proprietary datasets during preprocessing.

Neo orchestrates complex ML workflows (planning, training, deployment). Threats include tool misuse (e.g., executing arbitrary code via code execution environments used for feature engineering) and insecure tool integration with cloud providers.

Neo deploys and maintains models on scalable infrastructure. Threats include container/host compromise, privilege escalation via deployment credentials, and lateral movement within the organization's cloud environment.

Neo continuously monitors model performance. Threats include evaluation gaming (manipulating metrics to hide a compromised model) and drift/anomaly detection gaps if the monitoring system itself is compromised.

Not certain from the listing — there is no mention of identity access management, role-based access control, or regulatory compliance (e.g., EU AI Act, NIST) for the autonomous deployment of models.

Not certain from the listing — there is no explicit mention of Neo interacting with other agents or marketplaces, though cascading failures could occur if deployed models interact with other automated systems.