NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·May 27·8 min read
ChainCaps: Composition-Safe Tool-Using Agents via Monotonic Capability Attenuation
TL;DR
AI agents using external tools (like APIs or databases) risk unintended harm when composing multiple tool calls in sequence. ChainCaps introduces monotonic capability attenuation—a method that progressively restricts an agent's access permissions as it chains tool calls together.
✦ Why It Matters
Engineers can deploy multi-tool agents in production with formal safety guarantees against privilege escalation attacks.
Key Takeaways
- Implement ChainCaps to prevent permission laundering in agents.
- Focus on creating expert manifests for maximum security.
- Test your agents with ChainCaps to evaluate safety improvements.
- Understand the limitations of explicit-flow composition safety.
How It Works
ChainCaps operates by assigning a sink-specific capability budget to each value processed by the agent. As values are passed through different tools, their budgets are intersected, meaning they can only lose authority and cannot gain new permissions.
This mechanism ensures that even if an agent satisfies individual tool permissions, it cannot perform unauthorized actions through composition.