TensorFlow Lite for Microcontrollers (TFLM) is an embedded inference library targeting deeply resource-constrained bare-metal microcontrollers (typically with a single address space and no MMU) where every byte of Flash/ROM, every byte of RAM, and every CPU cycle matters.

Because TFLM is linked directly into device firmware as a local library rather than operating as a multi-tenant service or privilege boundary, its defensive programming model differs fundamentally from desktop, mobile, and server-side ML runtimes. Before reporting an issue or submitting a defensive-coding pull request, please read our Error Handling & Defensive Programming Guide.

1. The C++ API (Trusted): TFLM treats the firmware engineer calling the public C++ API as trusted.
2. The .tflite FlatBuffer Model (Trusted Structure): TFLM assumes that the .tflite model loaded into the interpreter is structurally valid and was produced by a trusted converter pipeline. To minimize binary footprint, TFLM intentionally omits a runtime FlatBuffer verifier as well as runtime bounds and integer-overflow checks on static model parameters. If a deployment receives models over an untrusted channel (such as OTA updates), the application firmware is responsible for authenticating and validating the integrity of the model before passing it to TFLM.

The following categories are working as intended under TFLM's design constraints and are not security vulnerabilities:

• Corrupted, mutated, or maliciously crafted .tflite FlatBuffer files: Crashes, aborts (TFLITE_CHECK / TFLITE_DCHECK), division-by-zero faults, integer overflows on static tensor dimensions, or out-of-bounds memory accesses that require modifying or crafting a .tflite model file (such as invalid FlatBuffer offsets, mismatched static tensor shapes, out-of-range operator attributes, or missing quantization parameters).
• Misuse of the C++ API: Passing invalid pointers, uninitialized buffers, or violating interpreter state-machine ordering from application code.
• Speculative or AI-scanned hardening reports: Automated static-analysis, fuzzer, or LLM-generated reports and pull requests that flag theoretical checks without a reproducible bug in a real TFLM deployment.