Scale computer-use 2.0 with open-source drivers, cross-OS sandboxes, and benchmarks for training, evaluation, and data generation.
Tool-call boundary fuzzing
Coverage-guided fuzz targets for the transport-free MCP tool-call boundary in
cua-driver-core: JSON-RPC request parsing, reserved-argument stripping,
argument normalisation, authorization, typed contract input parsing, and
ToolRegistry dispatch against stub tools. No platform adapter, display, or
input device is touched.
The target bodies live in cua_driver_testkit::boundary_fuzz; the files under
fuzz_targets/ are one-line libFuzzer wrappers. The same bodies run as bounded
smoke tests in cargo test -p cua-driver-testkit, so this directory is only
needed for real fuzzing.
The workspace pins a stable toolchain. cargo fuzz needs nightly only for
sanitizers, so run with --sanitizer none:
cd libs/cua-driver/rust
nix shell nixpkgs#cargo nixpkgs#rustc nixpkgs#cargo-fuzz --command \
cargo fuzz run --fuzz-dir fuzz --sanitizer none mcp_request \
fuzz/work/mcp_request fuzz/corpus/mcp_request -- -max_total_time=300
Targets: mcp_request, tool_arguments, typed_input_json, registry_invoke.
libFuzzer writes every new input it discovers into the first corpus directory
it is given. Passing an ignored fuzz/work/<target> directory first keeps
fuzz/corpus/<target> limited to the hand-picked seeds that are committed and
run by the smoke test. Omitting the directories makes cargo fuzz write into
fuzz/corpus/<target> directly; delete the generated entries before
committing if you do that.
Keep the CUA_DRIVER_* policy and permission-mode environment variables unset
while fuzzing so the process-default authorization path stays reproducible.
The Fuzz: Cua Driver tool-call boundary GitHub Actions workflow
(.github/workflows/fuzz-cua-driver.yml) runs every target for two minutes on
pull requests that touch the boundary crates, fifteen minutes nightly, and a
chosen budget on manual dispatch. A crash fails that target's job, prints the
panic and reproduce command in the job summary, and uploads the crash file and
log as an artifact.
A crash lands in artifacts/<target>/. Reproduce it with
cargo fuzz run --fuzz-dir fuzz --sanitizer none <target> artifacts/<target>/<file>,
fix the bug, then copy the file into corpus/<target>/ so the smoke test keeps
it as a regression case. Only hand-picked seeds are committed; libFuzzer's
generated corpus entries stay local.
See libs/cua-driver/docs/2026-09-04-tool-call-boundary-fuzzing-design.md.