Skip to content

Security: akios-ai/akios

Security

SECURITY.md

AKIOS Security Policy

Document Version: 1.4.4 Date: 2026-03-11

πŸ”’ Security Overview

AKIOS v1.4.4 is a minimal, open-source security cage for AI agents.
We take security very seriously β€” the entire product is built around hard containment, real-time protection, and provable audit.

This policy explains how we handle vulnerabilities in the open runtime.

πŸ“‹ Supported Versions

Version Supported Security Updates
1.0.x βœ… Active βœ… Full Support
<1.0 ❌ End of Life ❌ No Support

🚨 Reporting Vulnerabilities

DO NOT report security issues on public GitHub.

Send private reports to: security@akioud.ai

What to Include

  • Clear description of the vulnerability
  • Steps to reproduce
  • Affected versions
  • Potential impact (e.g. sandbox bypass, PII leak, cost overrun)
  • Suggested fix (if any)
  • Your contact info

Our Response Process

  1. Acknowledgment: Within 24 hours
  2. Triage & Validation: Within 72 hours
  3. Fix Development: 2–4 weeks (depending on severity)
  4. Coordinated Disclosure: We release fix + advisory together
  5. Credit: We publicly thank responsible reporters (Hall of Fame)

πŸ›‘οΈ What We Protect In v1.4.4

  • Security sandboxing (kernel-hard on native Linux, strong policy-based in Docker)
  • Syscall interception & resource quotas
  • Real-time PII redaction (44 patterns across 6 categories)
  • Enforced cost & infinite loop kill-switches
  • Merkle tamper-evident audit ledger
  • Non-root Docker container β€” containers run as UID 1001 by default
  • AST-safe condition evaluator β€” no eval() anywhere in the codebase
  • REST API β€” self-hosted FastAPI server (akios serve) with OpenAPI spec
  • Cage down data destruction β€” session data wipe (audit logs, workflow outputs)
  • HTTPS domain whitelist β€” selective network access for HTTP agent
  • --exec rejection β€” shell-injection trap blocks arbitrary command execution
  • akios http β€” secure HTTP requests with domain whitelisting & PII redaction
  • akios protect show-prompt β€” preview interpolated + redacted LLM prompts

Security Cage Lifecycle:

  • cage up β†’ activate protections β†’ workflows execute β†’ data generated
  • cage down β†’ session data destroyed (audit/, data/output/) β†’ input data preserved

Secure Data Erasure (cage down):

  • Each file is overwritten with cryptographically random bytes, fsynced to disk, overwritten with zeros, fsynced again, then deleted (unlink)
  • --passes N option repeats the overwrite cycle N times (default: 1)
  • --fast option skips overwrite and just deletes (for CI/CD cleanup where forensic recovery is not a concern)
  • ⚠️ SSD caveat: On solid-state drives with wear-leveling, overwritten sectors may be remapped. Extra passes have limited benefit. For maximum security on SSDs, use full-disk encryption (LUKS/FileVault) as the underlying layer.

Network Security:

  • Default: All network access blocked
  • allowed_domains whitelist for HTTP agent (specific domains only)
  • LLM APIs always permitted (OpenAI, Anthropic, Grok, Mistral, Gemini, Bedrock, Ollama)

No guarantees: No software is 100% secure.
Users must secure their environment and validate outputs.

πŸ“ž Contact

Security reports: security@akioud.ai
General questions: hello@akios.ai

Thank you for helping keep the cage strong.

AKIOS β€” Where AI meets unbreakable security
Use responsibly. Your safety and compliance are your responsibility. πŸ›‘οΈ

There aren’t any published security advisories