<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>LLM Agents | Hulin Wang</title><link>https://www.hulinwang.me/tag/llm-agents/</link><atom:link href="https://www.hulinwang.me/tag/llm-agents/index.xml" rel="self" type="application/rss+xml"/><description>LLM Agents</description><generator>Wowchemy (https://wowchemy.com)</generator><language>en-us</language><lastBuildDate>Sat, 01 Aug 2026 00:00:00 +0000</lastBuildDate><image><url>https://www.hulinwang.me/media/sharing.png</url><title>LLM Agents</title><link>https://www.hulinwang.me/tag/llm-agents/</link></image><item><title>Autonomous Vulnerability Patching Agent — DARPA AIxCC</title><link>https://www.hulinwang.me/project/aixcc-patching-agent/</link><pubDate>Fri, 08 Aug 2025 00:00:00 +0000</pubDate><guid>https://www.hulinwang.me/project/aixcc-patching-agent/</guid><description>&lt;p>&lt;strong>Role:&lt;/strong> Patching Component Lead, Team Shellphish (Aug 2023 – Aug 2025)&lt;/p>
&lt;p>The DARPA Artificial Intelligence Cyber Challenge (AIxCC) is a global competition for automated vulnerability discovery and repair in open-source and critical-infrastructure software. Team Shellphish was named one of 7 finalists in Aug 2024 (&lt;strong>$2M&lt;/strong> finalist award) and placed &lt;strong>5th&lt;/strong> in the final round in Aug 2025.&lt;/p>
&lt;h3 id="vulnerability-patching-agent">Vulnerability patching agent&lt;/h3>
&lt;ul>
&lt;li>Led development of a multi-agent vulnerability repair system integrating autonomous triage, patch generation, and validation across distributed services.&lt;/li>
&lt;li>Built a feedback loop combining fuzzing, proof-of-vulnerability re-tests, and deterministic regression tests, so that only patches that eliminate the vulnerability &lt;em>and&lt;/em> preserve functionality are accepted — &lt;strong>&amp;gt;95% patch correctness&lt;/strong> in the final scored rounds.&lt;/li>
&lt;li>Triaged and patched &lt;strong>30+ C/Java vulnerabilities&lt;/strong> in projects including curl, NGINX, and Apache ZooKeeper, covering buffer overflows, use-after-free, command injection, and path traversal.&lt;/li>
&lt;/ul>
&lt;h3 id="proof-of-concept-generation-agent">Proof-of-concept generation agent&lt;/h3>
&lt;ul>
&lt;li>Developed a multi-agent system that combines static-analysis results (CodeQL) and code context to generate inputs targeting specific vulnerabilities.&lt;/li>
&lt;li>Built a reachability-based program-slicing pipeline to focus agents on relevant code paths, using coverage-guided search to bypass blocking logic.&lt;/li>
&lt;li>Generated PoCs in &lt;strong>~5 minutes&lt;/strong> on average, triggering 10 vulnerabilities in ZooKeeper, Tika, and jsoup — &lt;strong>five of which Jazzer missed after 8 hours&lt;/strong> of fuzzing.&lt;/li>
&lt;/ul></description></item><item><title>Multi-Agent Triage for Static-Analysis Rules — Amazon</title><link>https://www.hulinwang.me/project/amazon-static-analysis-triage/</link><pubDate>Sat, 01 Aug 2026 00:00:00 +0000</pubDate><guid>https://www.hulinwang.me/project/amazon-static-analysis-triage/</guid><description>&lt;p>&lt;strong>Role:&lt;/strong> Applied Scientist Intern, Amazon, New York (May – Aug 2026)&lt;/p>
&lt;ul>
&lt;li>Reduced validation time for new static-analysis rules from &lt;strong>10–15 days to 2–3 days&lt;/strong> by automating triage of production findings at &lt;strong>90% accuracy&lt;/strong>, averaging 4 minutes and $0.50 per finding.&lt;/li>
&lt;li>Built a three-stage workflow with the Strands Agents SDK to summarize findings, gather code context, and verify results; scored &lt;strong>14 points higher than Claude Code&lt;/strong> using the same model.&lt;/li>
&lt;li>Supported all 47 rules from two internal analyzers across Python, Java, and JavaScript without per-rule or per-language tuning.&lt;/li>
&lt;li>Curated a benchmark of 600+ findings from one year of production triage records, covering every rule; achieved 95% verdict consistency across five runs.&lt;/li>
&lt;/ul></description></item><item><title>BuildBench: Benchmarking LLM Agents on Compiling Real-World Open-Source Software</title><link>https://www.hulinwang.me/publication/buildbench/</link><pubDate>Wed, 01 Apr 2026 00:00:00 +0000</pubDate><guid>https://www.hulinwang.me/publication/buildbench/</guid><description/></item></channel></rss>