---
source: arxiv
url: https://arxiv.org/abs/2606.18619v1
published_at: '2026-06-17T02:32:45'
authors:
- Zhengxiong Luo
- Mehtab Zafar
- Dylan Wolff
- Abhik Roychoudhury
topics:
- agentic-vulnerability-detection
- code-intelligence
- llm-security-agent
- fuzzing
- program-specification
- software-foundation-model
relevance_score: 0.9
run_id: materialize-outputs
language_code: en
---

# Code-Augur: Agentic Vulnerability Detection via Specification Inference

## Summary
Code-Augur is an LLM-agent security auditor that writes its assumptions as source-level assertions, then uses fuzzing to break those assertions. The paper claims this finds more vulnerabilities than current agentic bug finders and produced 22 new reports in open-source projects.

## Problem
- Agentic security tools can miss bugs when their hidden assumptions about safe code are wrong.
- A clean audit is hard to trust because dynamic checks usually test only suspected bugs, not the assumptions behind “secure” judgments.
- The issue matters for large open-source codebases where missed input-state mismatches can survive for years and lead to memory safety bugs.

## Approach
- Code-Augur first builds a threat model from the code, documentation, and build context, including attacker-controlled inputs, trust boundaries, and security-relevant state.
- When the LLM agent thinks a code site is safe, it records the reason as an executable invariant, such as an assertion that a pixel format channel count equals a color space channel count.
- A guided grey-box fuzzer runs the instrumented program and searches for inputs that falsify those assertions.
- If an assertion fails, Code-Augur triages the failure as either a real security issue or a bad invariant, then refines the invariant and repeats the loop.
- The final outputs are validated bug reports, a threat model, and surviving invariants that can be reused in later audits.

## Results
- On two benchmark sources, DARPA AIxCC and OSV, Code-Augur reports 34-370 more bugs than Claude Code and AIxCC-winning Atlantis, according to the excerpt.
- The benchmark evaluation used two frontier LLMs, showing the method can run on models such as Sonnet and DeepSeek rather than only curated security models such as Claude Mythos.
- Code-Augur found 22 new vulnerabilities in widely used open-source projects.
- Developers had fixed or confirmed 16 of those 22 new vulnerabilities at the time of the paper.
- The paper also reports several bug bounty rewards, but the excerpt does not give bounty counts or dollar amounts.
- In a gpsd case study, the inferred invariants helped identify an incomplete fix and related bug family across a four-month repair series.

## Link
- [https://arxiv.org/abs/2606.18619v1](https://arxiv.org/abs/2606.18619v1)