Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·Jun 9·5 min read
SecureClaw: Clawing Back Control of LLM Agents
TL;DR
Tool-using large language model (LLM) agents are vulnerable to unauthorized actions and sensitive data exposure. SecureClaw is a dual-boundary architecture that implements authorization at the action sink and confines plaintext data at the read boundary.
✦ Why It Matters
Engineers can implement SecureClaw to enhance the security of LLM agents against unauthorized actions and data exposure.
Key Takeaways
- Adopt SecureClaw for improved security in LLM applications.
- Utilize the PREVIEW→COMMIT protocol for state changes.
- Implement trusted gateways to handle sensitive data.
- Evaluate security measures using AgentDojo and ASB benchmarks.
How It Works
SecureClaw operates by establishing two security boundaries: one for authorizing actions at the effect sink and another for confining sensitive data at the read boundary. Sensitive reads are processed through a trusted gateway that replaces raw values with opaque handles, while writes that alter external states follow a PREVIEW→COMMIT protocol.
This ensures that only a trusted executor can commit changes that have been authorized, preventing unauthorized access to sensitive information.