Today we are announcing AKIOS, a new open-source runtime designed specifically for running untrusted AI agents on local machines.
The Problem: "curl | sh" for Agents
We are entering an era where developers routinely run AI agents that generate and execute code. Current practices often involve running these agents directly on the host machine or inside loosely configured Docker containers.
This is risky. Agents can hallucinate destructive commands, exfiltrate credentials, or accidentally modify critical system files. We need a runtime that assumes the agent is untrusted by default.
The typical "solutions" — Docker containers, VM isolation, or hope — leave critical gaps:
| Approach | Network Control | PII Redaction | Cost Kill-Switch | Audit Trail | Setup Time |
|---|---|---|---|---|---|
| Bare Metal | ❌ None | ❌ None | ❌ None | ❌ None | 0 sec |
| Docker | ⚠️ Manual iptables | ❌ None | ❌ None | ⚠️ Basic logs | ~30 min |
| VM (Firecracker) | ⚠️ NAT config | ❌ None | ❌ None | ⚠️ Syslog | ~1 hour |
| AKIOS | ✅ Policy-gated | ✅ 50+ patterns | ✅ Per-workflow | ✅ Merkle chain | 15 sec |
Enter AKIOS
AKIOS provides a strict sandbox based on seccomp-bpf and user namespaces. It follows a "deny-by-default" philosophy for both network and filesystem access.
Architecture Overview
graph LR
User["Developer / CI"] -->|"akios run workflow.yml"| CLI["AKIOS CLI"]
subgraph RUNTIME["AKIOS Security Runtime"]
CLI --> PolicyEngine["Policy Engine\n(validate + sign)"]
PolicyEngine --> Sandbox["Kernel Sandbox\nseccomp-bpf + cgroups"]
subgraph CAGE["Security Cage (ephemeral)"]
Sandbox --> FS["Filesystem Agent\npath-gated"]
Sandbox --> HTTP["HTTP Agent\nrate-limited"]
Sandbox --> LLM["LLM Agent\nbudget-gated"]
Sandbox --> Tool["Tool Executor\nallowlisted"]
end
FS --> PII["PII Redaction\n50+ patterns"]
HTTP --> PII
LLM --> PII
Tool --> PII
PII --> Audit["Merkle Audit\nSHA-256 chain"]
Audit --> KillSwitch["Kill-Switch\ncost + loop"]
end
KillSwitch -->|"sanitized output"| Output["Results\n(auditable)"]
Key Features
Strict Sandboxing — Agents run in isolated environments with minimal privileges. On Linux, this uses kernel-level controls (cgroups v2 + seccomp-bpf). On other platforms, Docker provides strong container isolation.
Explicit Policies — You define exactly which domains an agent can contact, which files it can read/write, and what commands it can execute. Everything else is blocked.
# Example: minimal policy for a doc-summary agent
version: 1
name: "doc-summary"
filesystem:
allow:
- path: "/workspace/docs"
mode: "r"
http:
allow:
- host: "api.openai.com"
methods: ["POST"]
rate_limit_per_min: 10
llm:
provider: "openai"
model: "gpt-4.1"
budget_usd: 0.50
audit:
merkle: true
pii_redaction: true
Real-Time PII Redaction — Over 50 patterns detected and redacted in under 50ms — SSNs, credit card numbers, API keys, emails, phone numbers — before they ever reach the agent.
Tamper-Evident Audit — Every action is cryptographically logged in a Merkle chain. If even one byte is altered, the chain breaks. This isn't just logging — it's cryptographic proof.
Cost Kill-Switches — Hard termination on budget exceed or infinite loops. Set a $1.00 limit per workflow and know it will be enforced.
How It Works: 30-Second Demo
# Install
pip install akios
# Initialize a project
akios init my-project
cd my-project
# Run a sandboxed workflow
akios run templates/hello-workflow.yml
What you'll see:
[akios] sandbox: seccomp-bpf, cgroups v2 (Linux)
[akios] pii: redaction enabled (50+ patterns)
[akios] audit: Merkle trail initialized
[akios] budget: $0.50 limit enforced
[akios] workflow: completed successfully in 1.2s
[akios] audit: chain verified ✓ (hash: a3b7c9...d2e1)
What Makes AKIOS Different
This isn't another AI framework. It's a security runtime.
- Frameworks give you tools to build agents. AKIOS gives you a cage to run them safely.
- Docker gives you container isolation. AKIOS gives you policy-gated, PII-redacted, cost-limited, cryptographically audited execution.
- Trust-me architectures rely on developers doing the right thing. AKIOS enforces it at the kernel level.
What's Next?
This is just the beginning. We are working on:
- Multi-agent orchestration — parallel execution with inter-agent policy gates
- Database connectors — PostgreSQL, SQLite with query-level policy enforcement
- Plugin system — community agents with sandboxed plugin architecture
- Hardware-accelerated inference — GPU access within the sandbox with memory isolation
Check out the Quickstart guide to give it a spin, or browse the source on GitHub.
Try It Yourself
pip install akios
akios init my-project
akios run templates/hello-workflow.yml
Secure your AI. Build with AKIOS.