Security
AI-Powered PDF Fuzzer Uncovers 31 Zero-Day Vulnerabilities in Major Readers
Researchers developed an LLM-driven approach to systematically probe PDF engines by generating complex API sequences, discovering critical flaws in Adobe, Foxit, and other popular tools.
A team of researchers has unveiled PDFuzzer, an automated fuzzing system that leverages large language models to identify security flaws in PDF readers. The tool combines LLM-assisted grammar generation with constraint solving to synthesize intricate sequences of API calls—moving beyond the simple, isolated test cases that traditional fuzzers rely on.
The Problem with Conventional Fuzzing#
Existing PDF fuzzer tools test individual API calls in isolation, which limits their effectiveness at discovering vulnerabilities. Many real-world security flaws only emerge when multiple API calls interact in specific sequences, meaning conventional approaches miss entire categories of bugs. This gap has left popular PDF readers inadequately tested against compound API call scenarios.
How PDFuzzer Works#
The system operates in stages. PDFuzzer begins by extracting specifications from JavaScript API documentation and analyzing execution traces to map relationships between individual calls. An LLM then constructs context-free grammars that capture the valid structure and dependencies of API sequences. Finally, a constraint solver generates concrete, meaningful API call chains for testing against the target PDF engines.
Findings Across Three Platforms#
Testing against Adobe Acrobat Reader, Foxit PDF Reader, and PDF-XChange Editor, PDFuzzer demonstrated substantial improvements over competing tools—achieving up to 48% better code coverage than established fuzzers like TypeOracle, Favocado, and Cooper, as well as other LLM-based approaches. The fuzzer discovered 31 previously unknown vulnerabilities ranging from information disclosure to arbitrary code execution, all of which were responsibly disclosed and resulted in bug bounties from vendors.
Validation and Impact#
Ablation studies confirmed that each component of the pipeline is necessary, with the LLM components achieving 93–98% accuracy across all stages. The work demonstrates how structured use of language models can significantly improve the effectiveness of security testing for complex software systems.
