Resources / FAQ
Frequently asked questions
Questions about securing AI, answered.
How Starseer secures the model supply chain, monitors agents at runtime, gives you observability into agents and MCP, and keeps policy under your control.
01
About Starseer
What is Starseer?
Starseer is an AI security platform that scans models before deployment, detects and contains agent and MCP threats at runtime, and enforces your own policy at the gateway. It is built on mechanistic interpretability.
Platform →What is mechanistic interpretability, and why does it matter for security?
Mechanistic interpretability studies how a model represents and processes information internally: what its layers, activations, and circuits do, not just what it outputs. Starseer uses it to detect backdoors, adversarial inputs, and unsafe intent that are invisible at the output level.
How is Starseer different from guardrails and red teaming?
Guardrails filter text going in and out, and red teaming probes outputs with prompts. Starseer reads what the model computes, so it can catch behavior no prompt in your test set triggers and intent that never shows up in the output.
02
Supply chain
Is an open-weight model safe to deploy?
Not without verification. A model with a trigger-conditioned backdoor scores exactly like a clean one on benchmarks, because the trigger is not in your test set. Model Scan inspects weights and activations before the model serves a request.
Open-weight models →Is the model I run the same one the vendor released?
Rarely. Teams fine-tune, quantize, distill, merge, or abliterate open models before shipping. Each change creates a new artifact, and the vendor’s evaluation no longer applies to it.
What does Model Scan check?
Structural fingerprinting against an approved baseline, backdoor detection through activation analysis and circuit tracing, and attested provenance. An approved model serves. A failed model is quarantined with a forensic report.
Model Scan →03
Runtime monitoring
What is AI runtime monitoring?
Watching what models and agents do in production, including prompts, tool calls, MCP requests, and model activations, and acting when behavior turns unsafe. Starseer detects drift, misalignment, and compromise, and contains a session in milliseconds.
Detection & Response →How is it different from EDR?
EDR watches processes, files, and network calls. An agent’s decision to reach for a credential is none of those. Starseer monitors the layer where agents decide and act.
What happens when a detection fires?
The session can be throttled or isolated before the action lands. The analyst gets a case file with the prompt, tool call, MCP request, application trace, and activation probe on one timeline, mapped to MITRE ATLAS.
What is AI detection engineering?
Applying security detection engineering to AI behavior: defining normal and unsafe behavior, writing detections, testing them on real traffic, and tuning them over time. Starseer writes detections as code, using YARA-X rules versioned in CI.
04
Observability
What does AI observability cover?
An inventory of every agent, model, MCP server, and identity, plus one timeline of every prompt, tool call, and MCP request, tied to the identity and workflow behind it.
Gateway & Observability →What is MCP security?
Securing the Model Context Protocol servers agents use to call tools: knowing which servers exist, which tools they expose, which identities call them, and what policy applies to each call.
Agent & MCP management →How does Starseer find agents and MCP servers?
Traffic through the gateway and AIDR passive mode is mapped per workflow to the agents, MCP servers, tools, and identities involved. MCP is in scope from the first week, with no application code changes.
What is observe mode?
Rules run on live traffic and record what they would have blocked, without blocking. Your team then marks each rule to block, tune, or retire, so nothing blocks until your traffic proves it should.
05
Control and deployment
Whose policy applies when I use a vendor’s model?
By default, the vendor’s. Starseer moves controls to the gateway, so your team writes the enforcement and it applies to every model equally, whether frontier, open-weight, or self-hosted.
Where can Starseer be deployed?
SaaS, on-premises, and air-gapped, from edge device to server rack with zero external calls. It drops into the request path in front of any endpoint without code changes.
Which frameworks does Starseer support?
Evidence is aligned to NIST AI RMF, ISO 42001, and the EU AI Act, and detections map to MITRE ATLAS.
06
Buying and partners
How is Starseer priced?
One annual subscription, priced in bands by monitored identities. Contact us for a quote.
Pricing →What does the Managed Service include?
Onboarding, 24x7 agentic observers on every session, MCP inventory and coverage, and analyst review on a monthly, weekly, or daily cadence, depending on the tier you choose (Audit, Monitor, or Manage).
Managed Service →Can MSSPs and resellers partner with Starseer?
Yes. Resellers, MSPs and MSSPs, and distributors get partner pricing and protected deal registration. MSSPs get one entitlement pooled across all client tenants.
Become a partner →Have a question we didn’t answer?
Talk to the team, or run the 4-minute diagnostic to see where your exposure is.