Scaled Cognition — agentic threat model

Uses a proprietary 'Agentic Pretrained Transformer' (APT) optimized for actions. Threats include adversarial prompt injection attempting to bypass the pre-trained policy constraints and model extraction/stealing of the closed-source APT architecture.

The system is explicitly stateless and retains no data, with personal data never leaving the enterprise. This significantly mitigates data exfiltration and knowledge-base poisoning risks, though transient data in transit must still be secured against interception.

Orchestrates actions across complex APIs and domains. While it enforces deterministic constraints and anti-hallucination rules when calling functions, threats include insecure tool integration and logic flaws in the deterministic constraint engine that could be exploited to trigger unauthorized API actions.

Offers Hosted or VPC/On-Premise deployment options. VPC and On-Premise deployments drastically reduce the external attack surface and mitigate lateral movement risks, though hosted options remain vulnerable to standard cloud infrastructure threats.

Features a built-in API simulation to test against mock APIs without hooking up backends. This allows for robust pre-deployment evaluation, though real-time drift and anomaly detection in production are not explicitly detailed.

Strong compliance posture highlighted by statelessness, local enterprise data residency, and deterministic execution constraints, aligning well with strict privacy regulations (e.g., GDPR, CCPA).

Not certain from the listing — The platform acts as a single system backing multiple channels (voice, chat, email), but there is no explicit mention of multi-agent collaboration, delegation, or marketplace interactions that would introduce cascading agent-to-agent trust risks.