AKIOS
Core Concepts (AKIOS V1.0)
Security-first runtime for AI agents: sandbox, PII redaction, audit, and cost/loop kills on every run.
Platform security levels
| Environment | Security level | Notes |
|---|---|---|
| Native Linux | Full (kernel-hard: cgroups v2 + seccomp-bpf) | Maximum isolation and enforcement |
| Docker (macOS/Windows/Linux) | Strong (policy-based) | Cross-platform; host FS perms/seccomp depend on host |
Docker is convenient and strong; use native Linux for the strictest guarantees (filesystem perms + seccomp on host).
Core protections
- Sandbox & quotas: syscall filtering, CPU/memory/file limits, default-deny network.
- PII redaction: >50 patterns on inputs/outputs.
- Audit: Merkle-style, tamper-evident logs.
- Cost/loop kills: budgets and token caps, loop detection.
Core agents (4)
- filesystem — read/write/stat within allowed paths.
- http — rate-limited web/API calls with redaction.
- llm — token/cost-tracked calls to OpenAI, Anthropic, Grok, Mistral, Gemini.
- tool_executor — allowlisted commands in a sandboxed subprocess.
Architecture
Single-process runtime, no sidecars. State is ephemeral per run; persistence is file-based (data/, audit/). Deployable as pip package or Docker (wrapper or direct). Minimal surface = easy to audit and operate in air‑gapped or high-assurance environments.