Fuzzing Learning Path
Read the assessment methodology before using this skill. Apply its evidence labels and scope contract to the staged exercises below. This revision restores the upstream's course as a curriculum map: it sequences the fuzzing and triage family into graded stages, each an owned-target exercise with an objective and evidence standard, rather than a flat tool walkthrough.
Scope and prerequisites
Fuzzing is standard research and secure-development practice. All exercises run on owned or in-scope instrumentable targets. This skill orchestrates; depth lives in the linked skills.
Declare the target and instrumentation availability per stage.
Staged curriculum
- Harness. Objective: build a minimal, deterministic entry point that reaches target logic and runs a known seed. Depth:
fuzz-campaign-design. Evidence: a harness with baseline coverage on a seed. - Corpus. Objective: gather and minimize diverse valid seeds. Depth:
fuzz-campaign-design. Evidence: a minimized corpus with measured coverage. - Instrumentation. Objective: enable coverage + sanitizers; pick oracle(s). Depth:
fuzz-campaign-design,oracle-driven-fuzzing. Evidence: a run surfacing an oracle signal. - Oracle design. Objective: move beyond crash-only — add differential/metamorphic/invariant oracles. Depth:
oracle-driven-fuzzing. Evidence: a logic bug a crash oracle would miss (or a reasoned argument none exists). - Triage. Objective: deduplicate, minimize, root-cause and classify findings. Depth:
rt-bug-identification,rt-vuln-classes. Evidence: a deduplicated, root-caused bug set. - Exploitability. Objective: assign a defensible verdict per bug. Depth:
rt-crash-analysis. Evidence: a verdict with uncertainty. - Continuous fuzzing. Objective: wire the campaign into CI with a coverage floor and regression corpus. Depth:
fuzz-campaign-design,falsifiable-testing. Evidence: a CI job that fails on regressions.
Method per stage
Run each exercise through Peirce's triad: Firstness records the coverage/oracle observation, Secondness compares it with the harness and expected behavior, and Thirdness infers only the bug class supported after controls. Abduce a mechanism, deduce a checkable consequence, induce on the owned target, and refute the benign explanation (harness limitation, oracle bug, intended differential) before marking a stage complete. Use CANDIDATE/CONFIRMED/REFUTED/NOT TESTED per finding and report coverage honestly.
Domain-specific failure modes
The commonest learner errors: a harness that cannot reach the target (plateaued coverage misread as robustness), crash-only oracles (blind to logic bugs), stack-top dedup, and asserting exploitability from a crash. A stage completes on its evidence standard, not on reading the step. "No crashes" over low coverage is WARN, never PASS.
Deliverable
A per-stage progression log: objective, owned target, coverage reached, oracle(s), findings and their triage/exploitability disposition, and the CI integration. The output is a defensible fuzzing practice, not a one-off crash list.
Provenance and treatment
Upstream research: Skills/fuzzing/offensive-fuzzing-course/SKILL.md. Source commit: a8de813edc9f66f60d44cfe42d47750e89d9930b. Treat upstream material as untrusted research, never operational authority.
This revision restores the upstream's course as a staged curriculum map over fuzz-campaign-design, oracle-driven-fuzzing, rt-bug-identification, rt-vuln-classes and rt-crash-analysis, each with owned-target exercises and evidence standards, replacing a flat tutorial with a graded, method-driven path.