Yu Feng
5 papers · Latest:
Options, Not Clicks: Lattice Refinement for Consent-Driven MCP Authorization
Conleash is a client-side middleware that uses a risk lattice and policy engine to provide consent-driven, boundary-scoped authorization for MCP tool invocations.
On the (In-)Security of the Shuffling Defense in the Transformer Secure Inference
This paper demonstrates a novel attack that bypasses the shuffling defense in Transformer secure inference, enabling model weight extraction.
Semia: Auditing Agent Skills via Constraint-Guided Representation Synthesis
Semia audits LLM agent skills by converting them into a Datalog fact base using CGRS, finding critical risks in over half of real-world skills.
Synthesizing Multi-Agent Harnesses for Vulnerability Discovery
AgentFlow synthesizes multi-agent harnesses to discover critical zero-day vulnerabilities in software like Google Chrome, outperforming existing methods.
Your Agent Is Mine: Measuring Malicious Intermediary Attacks on the LLM Supply Chain
This paper reveals critical security vulnerabilities in LLM API routers, demonstrating malicious code injection and secret exfiltration attacks.
📬 Weekly AI Paper Digest
Get the top 10 AI/ML arXiv papers from the week — summarized, scored, and delivered to your inbox every Monday.