Security Architecture
Hash: iron-sandbox
7 min read

Safety First
Isolating Intelligence

Autonomous agents need freedom to act, but within an ironclad sandbox. We don't just manage code; we manage the risk boundary of continuous evolution.

Safety First: Isolating Intelligence with SST Cover

/// 01The Problem of Agentic Overreach

In a world of GPT-5.4 reasoning, an agent given a simple task could technically expand its scope to re-architect your entire VPC. Without isolation, the "Eclawnomy" is just a high-speed accident waiting to happen.

CRITICAL_RISK_DETECTED

Shared infrastructure = Shared blast radius. We don't do that.

/// 02Account Vending: The Ultimate Sandbox

ClawMore uses an AWS Account Vending Machine. When you register a repository, we don't just create a folder; we provision a entirely separate, pristine AWS account.

  • Hard Boundaries: No cross-account traffic. Ever.
  • Ephemeral Access: Agents use short-lived STS tokens that expire the moment the mutation is verified.

/// 03SST v4: Code-Aware Security

By using SST v4 (our preferred framework for serverlessclaw), our security policies are literally woven into the infrastructure code. We define the "Minimum viable permission" for every agent role programmatically.

Frequently Asked Questions