Code assurance

Fuzzing

In reviewUpdated 2026-11-15Trust Center

In review. This record, and the tools and files it cites, are on a review branch. They are not in the main source tree yet, so you cannot check them in a copy of the main source today.

Status: as of 2026-11-11. The harnesses and seed corpora are committed and compile; the long fuzzing campaigns have not been run yet (see "Hours run" below). Nothing on this page claims coverage or findings that were not measured.

Targets

Two detached cargo-fuzz projects, each with a plain library of harness functions, libFuzzer binaries, committed seeds and a stable-Rust regression test that replays every seed and every crash reproducer.

targetprojectwhat it feedscode under test
png_decodefuzz/corearbitrary bytesaster_blueprint::png::decode_rgba (the image decoder)
font_tablesfuzz/corebytes as an OpenType filehudlife_opentype::font_tables::FontTables (hand-written table reader: cmap, hmtx, kern, GPOS)
opentype_shapefuzz/corebytes as a font filePlugin::register_owned + Shaper::shape (HarfRust backend)
bitmap_font_setfuzz/corebyteshudlife_core::font::parse_font_set
svg_importfuzz/coretextSVG path data and transform importers (aster-ui and hudlife-core)
graph_codecfuzz/coretext / bytesblueprint graph source and workspace file decoders
asset_formatsfuzz/corebytesWavefront OBJ mesh and .cube 3D LUT importers
appkit_wirefuzz/corebytessync wire Message::decode and document snapshot Doc::from_snapshot
appkit_textfuzz/coretextplugin manifest parser and inline Markdown
runtime_json_markdownfuzz/nativetextthe runtime's JSON parser and Markdown block/inline parsers
runtime_abifuzz/nativebytes split on 0xFF into six stringsthe C ABI (aster_app_new, set_view, set_state, click, type, set_theme, set_clock, advance, tree, state, free, with NULL arguments)

The harness fails on any panic. For the C ABI, which catches panics at the boundary and turns them into an error string, runtime_abi also fails when the error string says internal error (panic): containment is a safety net, not the goal. Decoders must return Err, never crash, never allocate without bound.

Aster decodes PNG only (aster-blueprint/src/png.rs); aster-blueprint/src/ui/image.rs states that there is no JPEG decoder in the tree and apps hand over pixels. Scene recordings and the network protocol are covered through appkit_wire. The network client in aster-appkit (sync.rs, TCP) feeds Message::decode, so that target is the one that sees untrusted network bytes.

How to run

Needs nightly Rust and cargo-fuzz (the image aster-unsafe-tools:L2 has both; tools/unsafe/Dockerfile.miri-fuzz builds it from the fleet image).

Shell
cd fuzz/core                       # or fuzz/native
cargo +nightly fuzz run --fuzz-dir . png_decode -- -max_total_time=1800 -rss_limit_mb=2048
cargo +nightly fuzz list --fuzz-dir .

Replay all seeds and reproducers on stable Rust, no libFuzzer needed (this is what CI should run on every change):

Shell
cd fuzz/core && cargo test
cd fuzz/native && cargo test

tools/unsafe/check.sh runs both replays and a short smoke of every target (default 20 s each, --fuzz-seconds N).

On a shared machine, run the fuzzers niced; the fuzz binaries read a corpus directory and write findings to -artifact_prefix, so keep both outside the source tree while fuzzing.

Seeds and reproducers

  • fuzz/<project>/corpus/<target>/ holds small seeds (a few hundred bytes to one 24 KB font); after a campaign the corpus is minimised with cargo fuzz cmin and only the minimised set is committed.
  • fuzz/<project>/regressions/<target>/ holds one file per crash found, committed together with the fix. The regression test replays it on every cargo test, so a fixed crash cannot come back unnoticed.

Hours run

libFuzzer campaigns completed0 hours
harness buildfuzz/core and fuzz/native compile; the libFuzzer binaries were not built
regression replayfuzz/core: passes on all committed seeds; fuzz/native: not completed

Why: the build machine's capacity was needed elsewhere, so the libFuzzer build and runs were stopped before they started. The target for the first campaign is at least 30 minutes of wall time per target (11 targets), niced, two at a time. This page is updated with hours, executions and findings when it is done.

Findings

None yet, because nothing has been fuzzed for long enough to say. Each finding will be fixed in its own minimal commit with a reproducer under regressions/. A fix that would change rendered pixels is not made; it is written up in the project notes instead.

OSS-Fuzz

Not applicable: OSS-Fuzz is for open-source projects and the runtime source is private. The local harness above stays an internal and CI tool.

Source docs/trust/FUZZING.mdlast changed 2026-11-15

More records

Every record in the Trust Center is public.