本文へ移動

Firewall rules

This page is rendered from the enforced policy itself (secure.cljk, guardrails kotoba-guardrails-2026-09-v2 / firewall kotoba-firewall-2026-09-v1). GET /v1/secure serves the same versions.

Guardrail templates

Evaluated deterministically on every request, both teams. block refuses with 403 guardrail-blocked before a job exists and leaves a durable receipt; redact masks before the model sees it; flag marks for review and never refuses.

A block rule is the AUP and has no switch. The redact / flag rules and each compliance framework can be turned on / off per signed-in account under Security controls (/secure); a rule an enabled framework requires is held on while that framework is on.

id action What
csam-block block CSAM references are refused on every request (AUP strictly-prohibited).
cbrn-block block CBRN/WMD uplift requests are refused on every request (AUP strictly-prohibited).
fraud-block block Fraud-as-a-service asks are refused (AUP strictly-prohibited).
secret-hygiene redact Live credential shapes in the prompt are masked before the model sees them.
pii-passkeys flag References to passkeys/private keys are flagged for review, never blocked.

Firewall: task × rung

task is the request's purpose (the bearer path defaults to code-review). Each row opens at its rung and above. The blue team has no rungs: the three standard tasks only. Observe mode is off (decisions are enforced).

task min rung none card identity business contracted
code-review identity
vulnerability-triage identity
remediation identity
payload-crafting contracted
c2-tooling contracted

payload-crafting and c2-tooling have an empty task list — even a contracted rung on the record does not open them; that is a separate operator decision. Refusals are 403 firewall-denied naming the tool.