AI NewsOpen sourceAnnouncement
GitHub open-sources a fuzzing agent that writes its own harnesses and triages every crash
GitHub Security Lab has released Taskflow Agent, an open source LLM-driven pipeline that picks C or C++ entrypoints, writes fuzzing harnesses, runs AFL++ and triages every crash into a report.
Image: GitHub Security Lab
Why it mattersA one-person security team can now point a repo at an autonomous fuzzer overnight and read a triaged report in the morning, so vulnerability hunting stops being gated on writing harnesses by hand.
Fuzzing is the quiet backbone of every modern security programme, and the reason most teams still get to it late is that writing a fuzz harness for someone else's C or C++ project is slow, unglamorous work. GitHub Security Lab put an autonomous agent in that seat on 24 September 2026, and open-sourced the whole pipeline.
The project is called the Taskflow Agent, and the fuzzing harness is the first workflow in a wider seclab-taskflows repository. It picks a C or C++ entrypoint from a target project, reads the build system, writes an AFL++ harness, runs the fuzzer, reads coverage reports, improves the harness when coverage plateaus, triages every crash, and writes a vulnerability report at the end. It runs unattended.
What is under the hood
Three layers do the work. A shell driver chains the pipeline stages. A set of YAML files, one per stage, holds the LLM prompts and the exit criteria. MCP tools give the model concrete hands and feet: compile a harness, run AFL, read a coverage file, save a crash. Claude Sonnet 5 is the default model, GitHub says on its post.
The fuzzing itself is structure-aware, so the model does not just throw random bytes at a parser. Custom mutators exist for JSON, XML, regex and a set of binary formats, and the pipeline builds a dictionary of tokens from code guards it sees during instrumentation, so inputs pass the shallow checks and reach the interesting code. The system does not use CodeQL, which places it beside GitHub's other AI security work rather than on top of it.
What the reader can do with it
The repository is public at github.com/GitHubSecurityLab/seclab-taskflows-fuzzing, under an open-source licence. GitHub's own guide runs a target through a Codespace, points the pipeline at a public repository, and reads the report. That works today for anyone with a GitHub account, without setting up AFL locally.
The post is direct about the risk: the agent executes arbitrary build commands from the target project on the host, without a container between the two, so the pipeline is only safe in a disposable environment with no elevated privileges. A Codespace, a throwaway virtual machine, or a container built for this purpose is the safe place; a developer laptop is not. GitHub does not publish CVE counts or a benchmark against alternatives, so treat the practical utility as an open question until independent runs are reported.
For a small application security team, or for a developer who owns a C or C++ dependency and does not want to write a harness by hand, this replaces a week of setup with an overnight run of somebody else's Codespace budget. The harness the agent writes is real code, kept in the run's artefacts, so a good one from an unattended run can be lifted into a repository's regular CI.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
