Internship/Placement - GNAT DAS: AI Automated Fuzz Testing Workflows

AdaCore: Helping Developers Build Software that Matters. Everything we do at AdaCore is centered around helping developers build safe, secure and reliable software. For 30 years, we have partnered with global leaders in aerospace.

AdaCore would like to understand better AI's benefits over more traditional forms of static analysis. Candidates for this position should have a strong interest in automated dynamic analysis verification and a good understanding of large language models (LLMs) and other AI components. A good working knowledge of Ada, C, and Python is essential. Experience with static parsing language tools like Libclang is highly beneficial.

Goals

The internship goal is to advance the state of the art in automated fuzz testing by integrating AI components into AdaCore's fuzz testing solution: GNATfuzz. This is then split into 2 phases (goal 1 and goal 2)—a research phase and an integration/implementation phase; the number of applicable activities for both goal 1 and goal 2 will depend on the length of the internship and the results of goal 1.

• Goal 1: Rapid Prototyping of AI integration and assessment, for example, we might:

• Investigate the benefits of corpus generation via AI.
• Perform a benchmark comparison of AI vs compiler-generated objects.
• Generation of interesting values for solving complex branch conditions, which are better at reaching complex code paths and therefore increasing coverage.
• Investigate the benefits of using AI to provide insights into a function signature for a C application.
• Identification of pointers to scalar objects and pointers to buffers.
• Grouping of buffer pointers and size components.
• Investigate the benefits of using AI to develop an advanced automated project fuzzing workflow.
• Understanding control flow graphs (statically generated via tools like CLANG).
• Understanding coverage reports.
• Using the information to dynamically identify the next best subprogram for fuzzing.
• Identification of subprograms with high complexity.
• Identification of components not applicable for fuzzing (blocking calls to external components, infinite main loops as seen in embedded systems, etc.).

• Goal 2: Targeted feature integration, for example, we might:

• Following the rapid prototyping/proof of concept phase, multiple directions can be taken depending on the results of goal 1 to integrate the most interesting/beneficial feature into the GNATfuzz product.
• AI Corpus Generation: Use an LLM to generate a diverse set of valid and boundary-case inputs for a target application based on its function signatures and documentation (if available).
• Intelligent Fuzzing Workflow Automation with AI: This project tackles the advanced workflow automation aspect of Goal 1. The goal is to move beyond fuzzing a single program and to create an intelligent system that can analyze large projects and decide what subprograms to fuzz and in what order.
• AI-Assisted Function Signature Analysis for C/C++: This project uses AI to interpret C/C++ function signatures to make fuzzing more effective. Fuzzing C/C++ code can be tricky because of pointers and complex data structures. An AI could help GNATfuzz understand these more clearly.

Skills required/nice to have:

• Experience/interest in fuzz testing or other forms of dynamic analysis and/or compilation (GCC or LLVM).
• Good programming skills in Python and C (or Ada).
• Experience/interest in language interfaces (i.e., libclang).
• Solid understanding and interest in LLMs and other primary aspects of AI.