OAuthLancer is a directed semantic fuzzer for OAuth 2.0 / OIDC authorization servers. It targets protocol-level logic violations rather than crashes: races in single-use credential consumption, redirect-boundary bypasses, PKCE downgrades, token-type confusion, header-parsing faults, scope escalation, outbound-fetch side channels, and revocation survival.
This artifact provides the system source code of OAuthLancer.
The system has four cooperating parts.
- Vulnerability pattern templates (VPT-1..8). Each class names one
violation signature, its flexibility-boundary anchor, a generation schema,
and an applicability condition.
implements/oauthlancer/vpt_map.pymaps protocol symbols and differential batteries onto the eight classes. - Template-directed two-layer generation. A skeleton layer assembles
protocol-conformant exchanges (login, consent, PKCE, token redemption)
through the target adapters in
protocol/; an anomaly layer injects class-specific deviations, including barrier-synchronized storms for the race templates (implements/oauthlancer/race_probes.py). - Graded class-specific oracles. Violations score in [0,1] per class
signature (
implements/oauthlancer/graded_oracle.py,core/security.py); scores feed the scheduler instead of acting as pass/fail only. - Conformance-scored scheduling. Dispatch priority combines class
frequency, severity, differential priors, and running score means. The
ablation switch (
implements/oauthlancer/ablation_fuzzer.py, arms A0-A5 plus SERIAL/NODIFF controls) isolates each contribution.
run_oauth_fuzzing.py CLI entry point
core/ engine: config, fuzzer, adapters, oracles, coverage, health
protocol/ per-target protocol adapters (13 authorization servers)
OAuthMapper/ SUT abstraction layer (symbol-level request/response)
configs/ target deployment configurations (13 servers)
targets_manager/ deployment managers and service configurations
implements/
oauthlancer/ VPT map, graded oracles, race probes, ablation arms, runner
scripts/ matrix runners and utilities
configs/ per-campaign lane configurations (coverage / ablation)
rq1_baselines/ coverage-instrumented Hydra provisioning and samplers
patches/ P1-P7 engine patches (idempotent; apply_patches.py)
tests/ smoke tests
validate/targets/hydra/poc/ mock login provider used by Hydra provisioning
Python 3.8+ (requests); Docker Engine 20.10+ for target provisioning;
Java 11+/Maven only for the optional CXF sample image build. The P1-P7
patches are already reflected in the packaged core/; apply_patches.py is
idempotent and included for provenance.
All scripts resolve the repository root by walking up from their own location
until core/fuzzer.py is found, so the package can be unpacked anywhere.
# unit test for the graded oracles (no Docker needed)
python3 implements/tests/test_graded_oracle.py
# arm dispatch smoke against a live target (default lane config: configs/oauth_logto.json)
python3 implements/tests/smoke_arms.py --target-config configs/oauth_logto.json --budget 300
# provision a target, then run one directed campaign (six-hour window default)
python3 targets_manager/cxf_oauth_manager.py --action start --config configs/oauth_cxf.json
python3 implements/oauthlancer/runner.py --target cxf_oauth --arm A0 --seed 1 \
--budget 10000 --base-config configs/oauth_cxf.json --run-dir runs/a0_s1
# ablation matrix: 6 arms x 3 targets x 5 seeds at an equal step budget
python3 implements/scripts/run_ablation_matrix.pyCampaign parameters follow the paper's protocol: generative arms run five
seeds in six-hour windows (--timeout-h, default 6; timeout_seconds: 21600
in the lane configs), and the ablation matrix holds an equal
10,000-protocol-step budget per run. A DONE marker records each run's
wall-clock budget; re-running a completed cell is idempotent.
MIT — see LICENSE.