GitHub Security Lab has released a new Fuzzing Taskflow — an autonomous, AI-powered fuzzing pipeline for C and C++ projects built on its Taskflow Agent framework — designed to take over the laborious parts of security testing with minimal human involvement.
Announced in a GitHub Blog post by Antonio Morales on September 24, the open-source tool lets a user point it at any GitHub repository using the familiar owner/repo format. From there, the agent works through the entire fuzzing workflow on its own: identifying suitable entry points, analysing the project's build system, writing fuzz harnesses, running the AFL++ fuzzer, reading coverage reports, improving the harnesses, triaging every crash, and finally writing a vulnerability report for each unique bug.
The Taskflow is powered by Claude Sonnet 5 by default, chosen after it passed GitHub Security Lab's internal testing without triggering guardrail issues. A different model can be configured through the project's model_config.yaml file.
The pipeline runs in a coverage-feedback loop. AFL executes in iterations with doubling time budgets — starting at 30 seconds and scaling up through 60, 120, 240 and 480 seconds to 960 seconds, or about 32 minutes per target. The fuzzer's queue is replayed against a separately built coverage-instrumented binary to obtain true source-line and branch coverage. Based on what it finds, the agent can add new seeds, edit harness sources, enrich AFL dictionaries, or skip cold error paths. It stops when coverage plateaus: two consecutive iterations gaining less than one per cent of absolute line coverage.
It also ships with four structure-aware fuzzing mechanisms: pre-built dictionaries and custom mutators for recognised formats such as JSON, XML, regular expressions, PNG and length-prefixed binary TLV data; a source-level dictionary mined from the target's own .c and .h files, including string literals and 32-bit numeric constants; a dynamically growing AFL dictionary enriched with tokens found near uncovered code; and a corpus-splice operator. An evolving corpus directory persists across iterations and campaigns, merged with afl-cmin so progress is never lost when a campaign restarts.
Crash triage is automated end to end. Crashes are minimised with afl-tmin, replayed under AddressSanitizer for stack traces, deduplicated by a stack-top hash, and previously known crashes are replayed to check whether upstream fixes resolved them. Each unique bug receives a markdown report with a verdict — vulnerability, library hardening issue, harness bug, out-of-memory, timeout, assertion failure or duplicate — along with root-cause analysis pointing to file and line, a reachability argument, an exploitability assessment, a suggested fix as a unified diff, and a regression-test sketch. The suggested patches are explicitly marked "review required," since the agent's analysis is limited by the model's understanding of the target.
The package also includes a live HTML dashboard showing a per-harness running pulse, coverage-trend tables with sparklines, a crash heatmap and an iteration timeline.
GitHub includes a pointed safety warning. Because the pipeline executes afl-fuzz, clang and arbitrary build commands chosen by the language model directly on the host — with no container in between — it must be run only in a disposable environment such as a Codespace or throwaway virtual machine, and without elevated privileges. A prompt-injected agent, the Lab notes, could otherwise do anything the user can.
Sources: "AI-powered fuzzing with the GitHub Security Lab Taskflow Agent" (The GitHub Blog, https://github.blog/security/applic...ith-the-github-security-lab-taskflow-agent/); "GitHub Security Lab launches AI-powered fuzzing agent for C/C++ projects" (TechGig, https://techgig.com/news/ai/github-...ered-fuzzing-agent-for-c/c-projects/134472293)
Announced in a GitHub Blog post by Antonio Morales on September 24, the open-source tool lets a user point it at any GitHub repository using the familiar owner/repo format. From there, the agent works through the entire fuzzing workflow on its own: identifying suitable entry points, analysing the project's build system, writing fuzz harnesses, running the AFL++ fuzzer, reading coverage reports, improving the harnesses, triaging every crash, and finally writing a vulnerability report for each unique bug.
The Taskflow is powered by Claude Sonnet 5 by default, chosen after it passed GitHub Security Lab's internal testing without triggering guardrail issues. A different model can be configured through the project's model_config.yaml file.
The pipeline runs in a coverage-feedback loop. AFL executes in iterations with doubling time budgets — starting at 30 seconds and scaling up through 60, 120, 240 and 480 seconds to 960 seconds, or about 32 minutes per target. The fuzzer's queue is replayed against a separately built coverage-instrumented binary to obtain true source-line and branch coverage. Based on what it finds, the agent can add new seeds, edit harness sources, enrich AFL dictionaries, or skip cold error paths. It stops when coverage plateaus: two consecutive iterations gaining less than one per cent of absolute line coverage.
It also ships with four structure-aware fuzzing mechanisms: pre-built dictionaries and custom mutators for recognised formats such as JSON, XML, regular expressions, PNG and length-prefixed binary TLV data; a source-level dictionary mined from the target's own .c and .h files, including string literals and 32-bit numeric constants; a dynamically growing AFL dictionary enriched with tokens found near uncovered code; and a corpus-splice operator. An evolving corpus directory persists across iterations and campaigns, merged with afl-cmin so progress is never lost when a campaign restarts.
Crash triage is automated end to end. Crashes are minimised with afl-tmin, replayed under AddressSanitizer for stack traces, deduplicated by a stack-top hash, and previously known crashes are replayed to check whether upstream fixes resolved them. Each unique bug receives a markdown report with a verdict — vulnerability, library hardening issue, harness bug, out-of-memory, timeout, assertion failure or duplicate — along with root-cause analysis pointing to file and line, a reachability argument, an exploitability assessment, a suggested fix as a unified diff, and a regression-test sketch. The suggested patches are explicitly marked "review required," since the agent's analysis is limited by the model's understanding of the target.
The package also includes a live HTML dashboard showing a per-harness running pulse, coverage-trend tables with sparklines, a crash heatmap and an iteration timeline.
GitHub includes a pointed safety warning. Because the pipeline executes afl-fuzz, clang and arbitrary build commands chosen by the language model directly on the host — with no container in between — it must be run only in a disposable environment such as a Codespace or throwaway virtual machine, and without elevated privileges. A prompt-injected agent, the Lab notes, could otherwise do anything the user can.
Sources: "AI-powered fuzzing with the GitHub Security Lab Taskflow Agent" (The GitHub Blog, https://github.blog/security/applic...ith-the-github-security-lab-taskflow-agent/); "GitHub Security Lab launches AI-powered fuzzing agent for C/C++ projects" (TechGig, https://techgig.com/news/ai/github-...ered-fuzzing-agent-for-c/c-projects/134472293)
