Files
Syoyo Fujita a8fb48fa91 Harden v3 C parser against untrusted glTF input
Threat model: parser is intended for server-side processing of attacker-
supplied glTF/GLB. Two adversarial review rounds plus a 1-hour libFuzzer
run (4 workers, ASan+UBSan, ~420M execs total, zero new artifacts) drove
this set of fixes. Concrete PoCs in tests/v3/security/ confirmed each
issue was exploitable on the prior code.

Path traversal (CRITICAL): tg3__load_external_file concatenated base_dir
with the JSON-supplied URI verbatim. A glTF with
"uri":"../../../tmp/secret" successfully loaded the file from outside
base_dir (verified by FNV64 match). New tg3__uri_is_safe rejects empty,
NUL, leading / or \\, Windows drive prefixes, and any '..' segment.
Path-buffer length checks switched to saturating subtraction so 32-bit
size_t cannot wrap.

Sign-coercion in byteStride: int32_t -1 was cast directly to uint32_t,
producing 0xFFFFFFFF and propagating into downstream count*stride math.
Restrict to glTF spec range: 0 (tightly packed) or [4, 252].

Index validation: parsed int32 index fields (accessor.bufferView,
primitive.indices/material/attributes, node.mesh/skin/camera/light,
scene.nodes[], skin.joints[], animation channel/sampler refs, MSFT_lod
ids, KHR_audio emitter/source refs, etc.) were stored unchecked. New
tg3__validate_indices walks every index field and returns
TG3_ERR_INVALID_INDEX on out-of-range. Gated by
tg3_parse_options.validate_indices, defaulting to 1.

Use-after-free on parse failure (PRE-EXISTING, surfaced by ASan during
fix verification): tg3_parse and tg3_parse_glb destroyed model->arena_
on error paths, but error messages on the user-facing tg3_error_stack
were arena-allocated. Any caller reading errors.entries[i].message
after parse failure read freed memory. tg3_model_free is now sole arena
owner; arena lives across error paths so messages stay valid until the
caller frees the model.

Other fixes:
- tg3_parse_glb: hoist tg3__model_init before header parse so callers
  can safely tg3_model_free on header failure.
- tg3__parse_primitive morph targets: when arena alloc returns NULL,
  pair with target_counts[ti]=0 so validators do not deref.
- Defensive 'if (!tarr) continue' in the morph-target validator loop.
- New Security Considerations block in tiny_gltf_v3.h documents the
  threat model, default-on validation, fs-callback contract, and error
  message lifetime.

Verification: 13 internal tests in tester_v3_c (incl. 7 new security
regressions covering path traversal absolute and relative, fs-callback
no-call assertion, byteStride wrap, OOB index, opt-in raw mode, ext
fields, and arena-message lifetime), 134/134 Khronos sample models
match v1 ground truth digest, 1-hour ASan+UBSan fuzz on the final code
clean.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-09 14:02:34 +09:00
..