Starseer is an AI security company that protects models and agents by looking at what a model is computing internally, not only at the text it produces. It is for organisations that run frontier, open-weight, fine-tuned or self-hosted models and need to know that a model is the one they approved and is behaving as expected.
The Starseer AI Oversight Platform applies that internal measurement across a model’s whole life, from the moment it is brought in to every request it serves. Its approach is summed up as “AI security built on interpretability”.
At A Glance
| Vendor | Starseer |
|---|---|
| Category | AI model security |
| Deployment | SaaS, on-premises Kubernetes or fully air-gapped |
| Agentic AI | Per-agent behavioural baselines, tool call anomaly detection, session containment |
| Certifications | SOC 2 Type I |
| Founded | 2025, Knoxville, Tennessee |
| Market classification | AI Model Security |
What Starseer Does
Some threats leave no trace in a model’s output. A backdoor may only trigger on a specific input, and a prompt injection or tool abuse can look normal in text until it reaches a real action. Testing what a model says will miss both.
Starseer works from the inside instead. Fingerprinting the model against an approved baseline catches what behavioural testing cannot, and the same baseline is then used to judge the model in production. Findings map to MITRE ATLAS, NIST AI RMF, ISO 42001 and the EU AI Act, so the evidence can go straight into an AI governance programme.
Products
AI Model Vulnerability Scan
The scan is for teams that take models from outside sources, such as open weights, fine-tunes, quantizations and community re-uploads, and need to know what they contain. It examines weights, activations and architecture rather than sampling outputs, and uses activation analysis and circuit tracing to surface trigger-conditioned behaviour and backdoors. It runs as a CI/CD deployment gate that marks each model approve, quarantine or investigate, keeps an audit-ready provenance record, and gives quarantined models a forensic report. Approved models become the baseline for runtime monitoring.
AI Detection & Response
Detection & Response is for security teams running models and agents in production. It builds a behavioural baseline for each model and each agent from activations and decision-to-action chains, and flags drift or swapped and tampered weights, so an agent cannot quietly be moved onto a different model. Detections fire on what an agent was computing and about to do, with any agent framework and any model endpoint. Rules are written in YARA-X and versioned like code. When something fires, it can throttle or isolate a single session or the whole workload within milliseconds, and it puts prompts, tool calls, traces and activation probes on one timeline for investigation.
AI Gateway
The gateway classifies every prompt by intent, complexity and domain before a model sees it and applies rules the customer writes. It stops injection, jailbreak and exfiltration attempts before inference, sends traffic only to approved endpoints such as Claude, GPT, Llama, Granite and open-weight models, and routes each task to a model of the right size. Each routing decision is logged as a trace.
How It Works
The platform sits in the request path in front of any model endpoint, with no application code changes and no SDK. It has no external dependencies, so it runs as SaaS, on on-premises Kubernetes, or fully air-gapped, including at the edge in drones, robotics, factory automation and medical devices.
A model passes through three checkpoints in order: the scan before deployment, the gateway on each incoming request, and runtime monitoring while it executes, with the approved baseline shared across all three. Output leaves the platform as SIEM-ready logs, API alerts that can be forwarded into SOAR, and OpenTelemetry traces.
Company Background
| Headquarters | Knoxville, Tennessee, United States. |
|---|---|
| Leadership | Tim Schulz, co-founder and CEO. |
| Funding | A $2 million seed round led by Gula Tech Adventures (2025). |
Note: This overview is compiled from Starseer’s publicly available information as of 3 November 2026. It is provided for information only. NHI Mgmt Group does not endorse any vendor, and nothing on this page represents the opinions of NHI Mgmt Group.