Перейти к содержанию

ADR-010: Security Model

Status

Accepted

Context

AI agents execute code in real systems, creating security risks. OWASP AST10 identifies critical threats for agentic systems.

Decision

Implement multi-layer security: 1. Isolation — Docker containers with seccomp, readonly filesystem 2. Secret management — HashiCorp Vault integration, short-lived credentials 3. Policy enforcement — OPA/Rego for access control 4. Audit logging — Hash-chained WAL, comprehensive logging 5. PII protection — Automated PII detection and redaction

Consequences

  • Positive: Secure agent execution in production
  • Positive: Compliance with security standards
  • Positive: Full auditability of agent actions
  • Negative: Increased complexity
  • Negative: Performance overhead for security checks
  • Governance: src/lib/43-governance.sh
  • Audit: src/lib/44-audit.sh
  • PII guard: src/lib/45-pii-guard.sh
  • Security docs: docs/architecture/security-model-2026.md