You can deploy this model and start earning money today!

drogba771/
circle-flatbuffers-index-validation-poc-DO-NOT-USE

Name: circle-flatbuffers-index-validation-poc-DO-NOT-USE
Availability: InStock

Views

No views yet

WARNING: SECURITY RESEARCH - PROOF OF CONCEPT ONLY

This repository contains malformed/crafted Circle (FlatBuffers) model files that trigger memory-safety crashes (out-of-bounds reads) in Samsung ONE's CircleImport.cpp loader. These are passive data files, not executable code — they do nothing on their own and only cause a crash when parsed by a vulnerable version of the Circle model loader.

DO NOT load these files with a production Circle/ONE runtime outside of an isolated, disposable research environment.

Findings Summary

File	Finding	CWE	Confirmation	Runtime Location
tensor_index_fuzz_poc.circle	Tensor index OOB — `subgraph.tensors[input_num]` via `op->inputs()[i]`	CWE-129	✅ Fuzz-confirmed (3 crashes)	CircleImport.cpp:855, :918, :936
buffer_opcode_manual_poc.circle	Buffer index OOB — `buffers[const_tensor.buffer]->data`	CWE-129	✅ Manual-confirmed	CircleImport.cpp:875
buffer_opcode_manual_poc.circle	Opcode index OOB — `operator_codes[opcode_index]` passed to ConvertSubgraph	CWE-129	✅ Manual-confirmed	CircleImport.cpp:1059
redzone_read_poc.circle	FlatBuffers relative offset reads past buffer end (ASan redzone)	CWE-125	✅ Fuzz-confirmed	operator field accessor → `op->inputs()`

Reproduction

Build Samsung ONE's circle-mlir toolchain with AddressSanitizer enabled
Load each .circle file through the standard model import path (CircleImport)
Each file passes VerifyModelBuffer() successfully but triggers an ASan-detected out-of-bounds read at the source locations listed above

All PoC files are structurally valid FlatBuffers — they pass the FlatBuffers verifier's structural checks. The vulnerability is that the verifier does NOT validate semantic consistency:

Tensor indices are not checked against tensors.size()
Buffer indices are not checked against buffers.size()
Opcode indices are not checked against operator_codes.size()

Root Cause

Samsung ONE's CircleImport.cpp uses unchecked std::vector::operator[] (or FlatBuffers Vector::Get()) to access tensors, buffers, and operator codes by their index, without verifying that the index is within bounds. The FlatBuffers verifier only ensures structural validity of the FlatBuffer fields, not that index values are valid references.

cpp

// CircleImport.cpp:855 — TENSOR INDEX OOB
auto &const_tensor = *subgraph.tensors[input_num];    // input_num from op->inputs()

// CircleImport.cpp:875 — BUFFER INDEX OOB
buffer = buffers[const_tensor.buffer]->data;           // tensor.buffer from FlatBuffer

// CircleImport.cpp:1059 — OPCODE INDEX OOB
auto func_or_error = ConvertSubgraph(..., model->operator_codes, ...);
// Inside ConvertSubgraph: opcodes[op->opcode_index] is used unchecked

Mitigation

Add bounds checks before array accesses in CircleImport.cpp:

cpp

if (input_num < 0 || input_num >= subgraph.tensors.size()) { /* error */ return; }
if (const_tensor.buffer >= buffers.size()) { /* error */ return; }
if (opcode_index >= operator_codes.size()) { /* error */ return; }

Alternatively, replace operator[] with .at() which throws std::out_of_range.

Files

tensor_index_fuzz_poc.circle (670 bytes) — verified by libFuzzer + ASan; triggers heap-buffer-overflow at tensor index access
buffer_opcode_manual_poc.circle (524 bytes) — manually crafted with tensor.buffer=999 and opcode_index=999
redzone_read_poc.circle (670 bytes) — fuzz-discovered; triggers SEGV via FlatBuffers offset past buffer end
report.md — Full vulnerability report

Credits

Discovered by Onurcan Genc via coverage-guided fuzzing with libFuzzer + ASan + UBSan.

drogba771/circle-flatbuffers-index-validation-poc-DO-NOT-USE