How can we trust our agents?
summary
Argues that static capability tracking in types makes agent safety model independent and compiler-enforced, not reliant on LLM alignment. Implements a prototype MCP server that validates Scala snippets with Scala 3 capability types, runs them in a vetted capability library and REPL as TACIT MCP server using a dual LLM design. Benchmarks show 100% prevention of leakage and injection attacks by design while utility slightly improves versus baseline; smaller models still generate useful code with compiler feedback.
AI-written summary · about
chapters (10)
- 0:00Trusting AI agents
- 1:27Risks from autonomous agents
- 3:12Agent safety approaches
- 5:18Agents as adversaries
- 6:42Capability safety in Scala 3
- 14:21Safety harnesses with capabilities
- 20:38Enforcing safety in execution
- 25:13Safety and expressiveness results
- 29:47Adoption and alternatives
- 32:58Key takeaways
related talks
40mSafe Coding with LLMs - Verifiable Scala Silver Bullet
19mGradient: Gradual Compartmentalization via Object Capabilities Tracked in Types
29mMCP Security: Keep Your AI Agents from Spilling the Tea
41mWhere Are We With Scala's Capabilities?
26mGovernance and Security of APIs and MCPs
01h 03mWhere Are We With Scala's Capabilities?
38mCoding Agents & Language Evolution: Navigating Uncharted Waters
51mAgentic AI and Model Context Protocol's Security Vulnerabilities