Show HN: Runtime security enforcement and capability scoping for agents
Researchers from Harvard and Carnegie Mellon have developed a runtime security system for AI agents that dynamically scopes sandbox permissions and monitors behavior to prevent attacks like payment fraud and prompt injection. The system has shown strong results on open benchmarks, and the team is seeking production data to further validate it before releasing a paper and open-source project.
Background
- The project comes from AI researchers at Harvard and Carnegie Mellon. They are developing a new approach to AI agent security — agents are autonomous programs that can use tools, access files, and make decisions over long sessions.
- The core problem: conventional "static sandboxing" (setting fixed limits at the start of a session) is vulnerable because agents operating over long periods can learn the boundaries and find ways to bypass them (e.g., splitting a payment to dodge a limit, or spreading malicious injections across a team of agents).
- Their proposed solution has two parts: (1) "dynamic capability scoping" — a sandbox that shrinks and shifts in real time, granting only the minimum permissions needed for each step rather than a fixed set for the whole session; (2) behavioral monitoring that flags suspicious patterns in actions that appear benign individually.
- The team has found these vulnerabilities across more than a dozen existing agent providers and frameworks but needs real production data to validate whether their system outperforms open benchmarks in practice. They are recruiting testers through a waitlist ahead of a planned paper and open-source release.