Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

Graph Is the Verifier: Agentic Reinforcement Learning for Interprocedural Vulnerability Detection

Source

arXiv

arxiv.org/abs/2607.26656v1

Read the source

Source

  • Source name: arXiv
  • Source kind: paper
  • Authors / organization: Yikun Li, Ting Zhang, Jiakun Liu, Jinfeng Jiang, Yuheng Yieh, Yixin Yang, Wen Bin Leow, Yide Yin, Yintong Huo, Eng Lieh Ouh, Lwin Khin Shar, David Lo
  • Venue: arXiv: cs.CR
  • Published: 2026-07-29
  • Collected: 2026-07-30
  • URL: http://arxiv.org/abs/2607.26656v1
  • PDF: https://arxiv.org/pdf/2607.26656v1
  • DOI: UNKNOWN

Plain-Language Brief

Researchers propose VulAgentRL, a framework for interprocedural vulnerability detection using agentic reinforcement learning (RL) and a Code Property Graph (CPG). The CPG verifies the evidence cited by the policy during training and inference. The policy is initialized by distilling teacher investigations. VulAgentRL outperforms state-of-the-art baselines on a strict pair-wise-correct metric while issuing fewer tool calls.

What the Document Says

  • The research focuses on interprocedural vulnerability detection using agentic reinforcement learning (RL).
  • A Code Property Graph (CPG) is used in the proposed VulAgentRL framework for both inference and training.
  • At inference time, the policy queries the CPG for callers, callees, dataflow, etc.
  • During training, the same CPG verifies the evidence cited by the policy.
  • Every node in the CPG carries a persistent integer identifier for exact comparison during verification.
  • The policy is initialized by distilling teacher investigations.
  • VulAgentRL outperforms state-of-the-art baselines, including frontier models, on a strict pair-wise-correct metric while issuing fewer tool calls.
  • The advantage of VulAgentRL persists on an out-of-distribution corpus and under class imbalance.

Key Concepts Mentioned

  • vulnerability detection
  • agentic reinforcement learning
  • Code Property Graph
  • verification
  • policy
  • distillation
  • strict pair-wise-correct metric

Claims Made by the Document

  • Agentic RL can close the gap in real-world vulnerabilities that require evidence from outside a function.
  • VulAgentRL outperforms state-of-the-art baselines on a strict pair-wise-correct metric while issuing fewer tool calls.
  • The advantage of VulAgentRL persists on an out-of-distribution corpus and under class imbalance.

Evidence Used

  • Results from a sample of real CVEs showing that 71.7% of vulnerable functions require evidence from outside the function.
  • Experimental results comparing VulAgentRL with state-of-the-art baselines.

Methods / Approach

  • Agentic reinforcement learning
  • Code Property Graph
  • Policy distillation

Limitations Stated by the Authors

The reward credits verdicts that are supported by evidence, but it is not stated whether false positives or negatives are considered.

Limitations Noted by the Agent

Agent interpretation — not from the source.

The text does not provide information on the complexity of the CPG or the computational resources required for VulAgentRL.

Open Questions

  • What are the false positive and false negative rates in VulAgentRL?
  • What is the computational complexity and resource requirements of the Code Property Graph?

Notes for Future Q&A

May include interpretation. Labeled where it does.

The text does not explicitly state whether the false positive and false negative rates are measured or discussed.

Raw Abstract / Excerpt

Real-world vulnerabilities often span multiple functions, yet most learning-based detectors classify each function in isolation: on a sample of real CVEs, we find that 71.7% of vulnerable functions require evidence from outside the function to be classified correctly. Agentic reinforcement learning (RL) could close this gap by enabling a model to gather that evidence itself, but it lacks a reliable reward, since a reward defined on the final verdict alone can be obtained without performing any investigation. We propose VulAgentRL, an agentic RL framework for interprocedural vulnerability detection built on a Code Property Graph (CPG). The CPG serves two roles: at inference time the policy queries it for callers, callees, dataflow, and other queries, and at training time the same graph verifies the evidence the policy cites. Because every CPG node carries a persistent integer identifier, this verification is an exact comparison rather than a textual match, so the reward credits verdicts that are supported by evidence. We further initialize the policy by distilling teacher investigations, and show that this warm start is necessary, since RL cannot acquire tool-use behavior it never samples. Under a repository-level split that prevents leakage, VulAgentRL outperforms state-of-the-art baselines, including frontier models, on the strict pair-wise-correct metric while issuing fewer tool calls, and its advantage persists on an out-of-distribution corpus and under class imbalance.