GitHub Security Lab launches autonomous AI fuzzing pipeline for C/C++ projects

GitHub Security Lab has released an autonomous fuzzing pipeline for C/C++ projects that replaces much of the manual work traditionally required in security testing. The Fuzzing Taskflow, built on top of the GitHub Security Lab Taskflow Agent framework, uses an LLM-driven agent to handle the entire fuzzing lifecycle without human supervision.

The tool, announced on September 24, 2026, targets a persistent problem in software security: even projects enrolled in OSS-Fuzz for years can still conceal critical bugs. The reason, according to the Security Lab team, is almost always the same — someone needs to monitor coverage, write new harnesses for unreached code, and triage crashes. The Fuzzing Taskflow aims to automate these steps.

How the Fuzzing Taskflow works

Users only need to point the pipeline at a GitHub repository. The agent then identifies suitable entry points, analyzes the build system, writes the harnesses, runs AFL++ (American Fuzzy Lop), reads coverage reports, improves the harnesses, triages every crash, and writes a vulnerability report for each unique bug. The entire process runs without a human babysitting it.

The simplest way to run the tool is to start a codespace from the repository at github.com/GitHubSecurityLab/seclab-taskflows-fuzzing and execute a single script: ./scripts/fuzzing/run_fuzzing.sh PROJECT. For example, pointing it at the popular XZ compression library (tukaani-project/xz) triggers the full pipeline automatically.

The pipeline is designed with a clean separation of responsibility: the LLM agent owns the decisions, and the MCP (Model Context Protocol) tools own the execution. The agent decides what to fuzz, what harness to write, and what coverage gap to chase next. The tools expose primitives like run_afl_for or compile_harness, but the agent never calls AFL or clang directly.

Coverage improvement loop and plateau detection

The core of the pipeline is an iterative coverage improvement loop. Each iteration, for each harness, the agent runs AFL for a time budget, replays the generated inputs against a coverage binary to get a real source-line and branch coverage report, and then reads the list of uncovered branches.

Based on what it finds, the agent chooses one of several actions: add a new harness, extend an existing one with a new input, or modify a harness to reach an uncovered branch. The time budgets double every iteration — from 30 seconds to 960 seconds (about 32 minutes per target) — so cheap early rounds capture low-hanging coverage while longer later rounds break through harder guard conditions.

The loop uses plateau detection to decide when to stop. Once two consecutive iterations each gain less than a configurable threshold (1% absolute line coverage by default), the agent determines it has hit diminishing returns and moves on. This prevents wasting compute time squeezing out the last fraction of a percent.

Structure-aware input generation

AFL’s default byte-level mutators work well on binary formats but struggle with structured, text-based inputs. The Fuzzing Taskflow ships four complementary mechanisms for producing structure-aware inputs.

For targets whose input format is recognised — JSON, XML, regex, PNG, or length-prefixed binary TLV — the pipeline includes pre-built AFL dictionaries and custom mutator C files. The JSON mutator does token splicing and balanced-bracket duplication; the XML one knows about tags, entities, and billion-laughs tokens; the regex one carries real ReDoS patterns. Each mutator delegates half of its mutations back to AFL’s default byte mutator to preserve the engine’s randomness.

For formats the pipeline does not recognise, it generates a custom mutator on the fly by scanning the target’s own C and header files. It extracts string literals and 32-bit numeric constants from defines, case statements, and enums, and filters out everything else.

Datos clave

Aspect	Detail
Developer	GitHub Security Lab
Target	C/C++ projects on GitHub
Core technology	LLM agent (Claude Sonnet 5 by default) + AFL++
Key innovation	Autonomous harness writing, coverage improvement, crash triage, and report generation

Safety and model selection

The taskflow runs AFL, Clang, and arbitrary build commands chosen by the LLM directly on the host, with no container in between. A prompt-injected agent could, in principle, do anything the user can. GitHub recommends running it only inside a disposable environment such as a Codespace or a throwaway VM, without elevated privileges.

By default, the pipeline uses Claude Sonnet 5 because it passed all internal tests without issues. Users can select a different model by modifying the model configuration file.

For Indian developers and security teams working on open source C/C++ projects, this tool could reduce the manual effort required to maintain fuzzing coverage. Projects that lack dedicated security engineering resources — common among Indian startups and small teams — may benefit from the autonomous pipeline’s ability to write harnesses and triage crashes without requiring deep fuzzing expertise.

Source: GitHub Blog AI — AI-powered fuzzing with the GitHub Security Lab Taskflow Agent (https://github.blog/security/application-security/ai-powered-fuzzing-with-the-github-security-lab-taskflow-agent/)