Skip to content

Model routing

Hacker Sidekick members do two very different kinds of work: everyday engineering and documentation, and offensive-security tasks that a conservative model may refuse. Your instance can keep those apart by sending each request to one of two tiers:

  • the default tier, served by your default-tier model, and
  • the sensitive tier, served by your sensitive-tier model (one suited to exploitation, payload and red-team work).

Members don’t choose. Desktop asks for its usual model names, and your instance decides.

  1. Desktop sends a chat request to your instance.
  2. Your instance looks at the routing policy:
    • always default or disabled: the request goes to the default-tier model.
    • always sensitive: the request goes to the sensitive-tier model.
    • inherit: the classifier model reads the member’s latest messages and picks default or sensitive, and the request goes to that tier’s model.
  3. The chosen model answers, streaming back to Desktop.

With the inherit policy, every request first goes to your classifier model, which reads the member’s last two messages and answers default or sensitive.

It routes to sensitive for work such as penetration testing, exploitation and proofs of concept, payloads and evasion, attack chains and lateral movement, credential attacks, malware analysis, and active or intrusive scanning. It routes to default only for clearly non-offensive work: general software development, documentation, defensive-only guidance, and high-level concepts. When it’s unsure, it picks sensitive.

If the classifier is missing, can’t be reached, or returns something unreadable, the request goes to the default tier rather than failing. That’s deliberate: a confused or manipulated classifier can only narrow what a member reaches, never widen it. Repeated identical messages from the same member reuse the earlier decision for 10 minutes, so the classifier isn’t asked twice.

Hacker Sidekick sets the routing policy, and which model serves each tier, when your instance is provisioned. To change either, contact Hacker Sidekick.

Policy Behavior Classifier runs?
inherit The classifier decides per request. The default for new instances. Yes
always default Every request uses the default-tier model. No
always sensitive Every request uses the sensitive-tier model. No
disabled Same as always default. No

When to use which:

  • inherit is the normal choice when you have two different models and want each used for what it’s good at.
  • always sensitive suits teams that do offensive work nearly all day, or when you only trust one model with your data anyway.
  • always default or disabled suits teams that should never receive offensive-security assistance.

Your instance always answers as long as a default-tier model is enabled:

If this is missing or disabled… …this is used instead
Sensitive-tier model Default-tier model
Classifier model Default-tier model
Embedding model Nothing. Workspace indexing in Desktop reports that embeddings aren’t configured.

Audit & Usage shows how many requests went to each tier, per day, so you can check the classifier is behaving the way you expect.