PraxisAI — agentic threat model

Not certain from the listing — The specific deep learning and reasoning models used are not disclosed. Potential threats include adversarial examples targeting sensor data inputs, model reprogramming, and misaligned outputs that could lead to false negatives in critical machine failure predictions.

Integrates structured and unstructured factory data alongside real-time machine data. This introduces significant risks of data poisoning (corrupting sensor streams to mask failures) and data exfiltration of sensitive manufacturing processes or proprietary engineering manuals.

Supports custom workflow creation and reasoning for critical machinery. Vulnerabilities here include insecure tool integration (e.g., direct connections to SCADA/PLC systems) and tool misuse, where an attacker could trigger unauthorized physical actions or machine shutdowns.

Not certain from the listing — The hosting, sandboxing, and network isolation details are unspecified. In industrial deployments, a lack of strict network segmentation between the AI platform and the Operational Technology (OT) network could allow lateral movement and host compromise.

Not certain from the listing — While the platform predicts and prevents failures, the internal observability, logging, and guardrail mechanisms are not detailed. Gaps in drift detection could cause the model to miss critical anomalies as machinery ages or undergoes maintenance.

Not certain from the listing — No compliance certifications (such as SOC2, ISO 27001) or specific identity and access management (IAM) controls are mentioned. Weak authorization controls could allow unauthorized users to modify critical machinery workflows.

The platform builds manufacturing-specific AI agents to handle complex situations. This multi-agent potential introduces risks of cascading failures, where a compromised or malfunctioning agent propagates incorrect state information to other agents in the workflow.