PBFuzz: Agentic Directed Fuzzing for PoV Generation

2026-06-30 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy, Software Development & Engineering · Depth: Expert, extended

Summary

PBFuzz is an agentic directed fuzzing framework designed for efficient Proof-of-Vulnerability (PoV) input generation in software security. It addresses the limitations of traditional directed greybox fuzzing and LLM-assisted methods by mimicking human experts. PBFuzz iteratively analyzes code to extract semantic reachability and triggering constraints, hypothesizes PoV plans, encodes them as test inputs, and refines its understanding using debugging feedback. Key features include autonomous code reasoning, custom MCP tools for on-demand program analysis, persistent memory to prevent hypothesis drift, and property-based testing for efficient constraint solving. On the Magma benchmark, PBFuzz triggered 57 vulnerabilities, exclusively finding 17, within a 30-minute budget per target, demonstrating a 25.6x efficiency gain over AFL++ with CmpLog, at an average API cost of \$1.83 per vulnerability.

Key takeaway

For AI Security Engineers or vulnerability researchers tasked with generating Proof-of-Vulnerability (PoV) inputs, especially for complex software, you should consider adopting agentic, property-based fuzzing frameworks like PBFuzz. This approach significantly accelerates PoV discovery and improves consistency compared to traditional or basic LLM-assisted fuzzers, particularly for vulnerabilities requiring intricate structural invariants. However, be mindful of its limitations regarding systemic scope and construct validity bias, which may necessitate hybrid strategies for certain bug types.

Key insights

PBFuzz leverages agentic LLMs and property-based testing to bridge semantic understanding and efficient PoV input generation.

Principles

• PoV generation is finding counterexamples violating safety properties.
• Agentic LLMs can reason, plan, orchestrate tools, and self-reflect.
• Property-based testing systematically explores input spaces, preserving validity.

Method

PBFuzz uses a four-phase workflow: PLAN (infer constraints), IMPLEMENT (encode to generators), EXECUTE (PBT fuzzing), and REFLECT (diagnose/refine hypotheses).

In practice

• Use agentic LLMs for dynamic context search and hypothesis validation.
• Employ persistent memory to prevent hypothesis drift.
• Synthesize parameterized input generators for efficient constraint solving.

Topics

• Agentic AI
• Directed Fuzzing
• Proof-of-Vulnerability
• Property-Based Testing
• Software Security
• Vulnerability Discovery

Code references

• R-Fuzz/magma

Best for: Research Scientist, AI Scientist, AI Security Engineer, Software Engineer

Related on AIssential

• Code-Augur: Agentic Vulnerability Detection via Specification Inference — Agentic vulnerability detection improves by making LLM assumptions explicit and continuously validating them with fuzzing.
• DIG: Oracle-Guided Directed Input Generation for One-Day Vulnerabilities — Patch analysis with LLMs can synthesize explicit vulnerability preconditions, guiding effective PoC generation.
• Mutation testing for the agentic era — Mutation testing tools MuTON and mewt, powered by Tree-sitter and AI agents, enhance software quality by efficiently identifying untested code paths.
• Capture the Flags: Family-Based Evaluation of Agentic LLMs via Semantics-Preserving Transformations — CTF challenge families and Evolve-CTF enable robust evaluation of agentic LLMs against code transformations.
• FuzzingRL: Reinforcement Fuzz-Testing for Revealing VLM Failures — FuzzingRL automatically generates adversarial questions to expose and amplify Vision-Language Model vulnerabilities through iterative refinement.
• Introducing the Red Agent POV Series — AI-powered offensive security agents can autonomously discover complex, logic-driven vulnerabilities at scale, surpassing traditional methods.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.