Luncat ka eusi
Aman Guardrails

Kontrol kaamanan

Kontrol runtime tina lingkungan panalungtikan — aturan anu dipertontonkeun di dieu nyaéta aturan anu leres-leres dievaluasi ku pihak anu ngajinkeun aksés.

Guardrails

kotoba-guardrails-2026-09-v2

Kawijakan eusi anu dilarapkeun kana badan (body) unggal request. Kaputusan sipatna deterministik; blok bakal nolak job saméméh aya kuota nu dipaké.

Ngaran Aksi Pedaran Pola Aktif
csam-block Blokir CSAM references are refused on every request (AUP strictly-prohibited). 0 Salawasna hurung (AUP)
cbrn-block Blokir CBRN/WMD uplift requests are refused on every request (AUP strictly-prohibited). 0 Salawasna hurung (AUP)
fraud-block Blokir Fraud-as-a-service asks are refused (AUP strictly-prohibited). 0 Salawasna hurung (AUP)
secret-hygiene Samarkeun Live credential shapes in the prompt are masked before the model sees them. 3
pii-passkeys Tandaan References to passkeys/private keys are flagged for review, never blocked. 2

Firewall

kotoba-firewall-2026-09-v1

Tool tugas anu bisa dipaké ku agent, jeung tahap kapercayaan anu diperlukeun ku unggal tool. Panelepon anu handapeun tahapna bakal ditolak (mode observasi ngan ukur ngarékamna).

Tool Tahapan anu diperyogikeun Tugas anu kabuka
code-review Identitas geus diverifikasi code-review
vulnerability-triage Identitas geus diverifikasi vulnerability-triage
remediation Identitas geus diverifikasi remediation
payload-crafting Dikontrak — (ditolak sacara standar)
c2-tooling Dikontrak — (ditolak sacara standar)

Kapatuhan

kotoba-compliance-2026-09-v1

Coverage = automated runtime controls only — not a certification.

Framework Wilayah Kontrol otomatis Aktif
APPI JP prompt-pii-redaction audit-log-retention screening-evidence-receipts
PCI DSS 4.0 Global card-data-never-stored secret-hygiene-redaction audit-log-retention
OWASP LLM Top-10 Global guardrails-csam-cbrn firewall-task-deny screen-model-block-receipts
NIST AI RMF US session-projection-uncalibrated assurance-ladder-evidence
Panyimpenan log audit 180 poé
Lokasi data Teu ditangtukeun
Retensi data nol Dirarancang pikeun ZDR (diaktifkeun dina kontrak)