Model artifacts cross three distinct trust boundaries on the way to execution. Each public advisory below maps the serialized field that enters the loader, the missing invariant at that boundary, and the witnessed failure mode.
The index loader allocates its level-0 buffer from serialized max_elements_ and then reads cur_element_count records into it without checking that count fits capacity.
A negative signed pickle storage_offset sign-extends to uint64_t and wraps storage_offset + nbytes modulo 2^64 past zero, bypassing the archive end-of-buffer check and reading before the allocation.
Repeating a tensor name in an .imatrix file skips an empty()-gated vector resize on the second entry while writing the larger second byte count into the smaller retained buffer.
Convolution shape inference defaults dilations[] from input rank while iterating over weight spatial rank, reading 8 bytes past the end of dilations[] during session load.
CUDA EXL3 reconstruct dispatch · Out-of-range function dispatch
Checkpoint-derived quantization bitwidth K indexes a fixed 24-entry CUDA function-pointer array without checking the 1..8 domain, calling past the dispatch table.
Quantized checkpoint group indices (g_idx) are used directly as write offsets during sequential weight permutation without checking against the allocated group count.
Out-of-range quantization group indices in a GPTQ checkpoint reach the Triton dequantization kernel unchecked and index past scales and qzeros on GPU memory.
Constructing a tensor from raw serialized bytes when declared shape product disagrees with payload byte length triggers an unhandled panic during model load.
Deserializing a remote GRAPH_COMPUTE graph reconstructs custom-operation callback addresses from untrusted wire bytes and invokes them during execution.
Remote tensor descriptors with quantized row dimensions that violate block-alignment rules reach GGML_ASSERT instead of returning an RPC validation error.
Crafted regex patterns inside structured-output JSON schemas leave the converter state empty and dereference a null pointer during grammar compilation.
Hardening patches and public advisories landed upstream across model graph compilers, loaders, and codec libraries, grouped by ecosystem. Findings still under coordinated embargo are omitted until upstream fixes ship.
01
TensorFlow Core & Grappler
Graph import, variant decoding, and optimizer pass bounds merged into tensorflow/tensorflow.
Reject [gvar] glyph variation data offsets that wrap past 2^32.
B1Integer wraparound
Coordinated disclosure inquiries follow the Crucible disclosure policy or [email protected]. Active embargoes and unmerged reports stay off this page until maintainers release a fix.