Home/Advisories

Advisories

Public vulnerability disclosures and merged upstream patches across model parsers, loaders, and runtimes.

Where validation breaks

Three parser-to-kernel boundaries

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.

Distribution of 20 public CVEs across container framing, schema and graph construction, and native kernel dispatch Boundary 1 Container and Framing holds 8 CVEs; Boundary 2 Schema and Graph holds 7 CVEs; Boundary 3 Kernel and Dispatch holds 5 CVEs. BOUNDARY 1 Container & Framing Offsets, header counts, string & wire framing 8 public CVEs BOUNDARY 2 Schema & Graph Rank agreement, vocab IDs & node topology 7 public CVEs BOUNDARY 3 Kernel & Dispatch Quant tables, dispatch domains & media/GPU 5 public CVEs FAILURE MODES (20 PUBLIC CVES) Out-of-bounds read / write 7/20 Null / assertion / recursion 8/20 Dispatch / deserialization 3/20 Integer wrap / underflow 2/20
Disclosed vulnerabilities

Public CVE catalog

20 public disclosures

CVE-2026-124959 CERT/CC VU#681158 (KB note pending)
B1 · Container & Framing

nmslib/hnswlib & spotify/voyager

HierarchicalNSW::loadIndex · Heap out-of-bounds write

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.

CVE-2026-115571 GHSA-cfhc-rw9j-rp26
B1 · Container & Framing

leejet/stable-diffusion.cpp

Pytorch Zip/Pickle tensor loader · Integer wrap / heap OOB read

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.

CVE-2026-70659 GHSA-6q7q-2v9c-x8m7
B1 · Container & Framing

leejet/stable-diffusion.cpp

Quantization imatrix loader · Heap out-of-bounds write

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.

CVE-2026-14647 VDB-376160
B2 · Schema & Graph

onnx/onnx

convPoolShapeInference · Heap out-of-bounds read

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.

CVE-2025-66455 GHSA-2vh9-42vm-xmv2 · co-credited
B1 · Container & Framing

InternLM/lmdeploy

DistServe ZMQ message plane · Untrusted deserialization

Unauthenticated Distributed Serving messages reach PyZMQ recv_pyobj pickle deserialization before peer trust is established. Fixed in LMDeploy 0.16.0.

CVE-2026-84286 CERT/CC VU#369611
B3 · Kernel & Dispatch

turboderp-org/exllamav3

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.

CVE-2026-116502 VDB-426401 · assigner-published
B3 · Kernel & Dispatch

turboderp-org/exllamav2

QMatrix::make_sequential · Heap out-of-bounds write

Quantized checkpoint group indices (g_idx) are used directly as write offsets during sequential weight permutation without checking against the allocated group count.

CVE-2026-86288 VDB-399447
B3 · Kernel & Dispatch

ModelCloud/GPTQModel

Triton dequantization kernel · Device out-of-bounds read

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.

CVE-2026-86289 VDB-399448
B1 · Container & Framing

ollama/ollama

readGGUFV1String · Integer underflow / panic

The GGUF v1 string reader subtracts 1 to strip a trailing null byte without checking for zero length, panicking inside bytes.Buffer.Truncate(-1).

CVE-2026-75090 VDB-391327
B2 · Schema & Graph

EricLBuehler/mistral.rs

GGUF chat-template vocabulary loader · Out-of-range index / panic

GGUF metadata token IDs (BOS, EOS, UNK) index the loaded vocabulary array before verifying that each identifier falls inside the token table.

CVE-2026-75093 VDB-391335
B2 · Schema & Graph

sonos/tract

Tensor::from_raw_dt_align · Shape-extent mismatch panic

Constructing a tensor from raw serialized bytes when declared shape product disagrees with payload byte length triggers an unhandled panic during model load.

CVE-2026-78147 VDB-394534
B3 · Kernel & Dispatch

ggml-org/llama.cpp (ggml-rpc)

RPC GRAPH_COMPUTE custom-op · Controlled indirect call

Deserializing a remote GRAPH_COMPUTE graph reconstructs custom-operation callback addresses from untrusted wire bytes and invokes them during execution.

CVE-2026-78148 VDB-394535
B2 · Schema & Graph

ggml-org/llama.cpp (ggml-rpc)

RPC ggml_graph_plan · Null dereference

A serialized top-level node ID of 0 resolves to a null sentinel pointer and reaches ggml_graph_plan without a non-null node check.

CVE-2026-86317 VDB-399508
B2 · Schema & Graph

ggml-org/llama.cpp (ggml-rpc)

RPC tensor reconstruction · Assertion abort

Remote tensor descriptors with quantized row dimensions that violate block-alignment rules reach GGML_ASSERT instead of returning an RPC validation error.

CVE-2026-17512 VDB-383382
B3 · Kernel & Dispatch

ggml-org/whisper.cpp

log_mel_spectrogram padding · Heap out-of-bounds read

Reflective audio padding assumes at least FFT_SIZE / 2 + 1 input samples; shorter clips step std::reverse_copy past the end of the sample buffer.

CVE-2026-10298 VDB-367591
B2 · Schema & Graph

ggml-org/whisper.cpp

whisper_model_load · Null dereference / abort

Zero-valued model hyperparameters allocate zero-sized backing buffers and trigger null-pointer copies or assertions while loading model weights.

CVE-2026-17513 VDB-383383
B2 · Schema & Graph

ggml-org/whisper.cpp

ggml_ftype_to_ggml_type · Reachable assertion

An unsupported file-type enum in a Whisper checkpoint header hits an unconditional GGML_ABORT instead of failing model load cleanly.

CVE-2026-17500 VDB-383353
B1 · Container & Framing

ggml-org/llama.cpp

JSON-schema regex converter · Null dereference

Crafted regex patterns inside structured-output JSON schemas leave the converter state empty and dereference a null pointer during grammar compilation.

CVE-2026-17501 VDB-383354
B1 · Container & Framing

ggml-org/llama.cpp

JSON-schema-to-GBNF converter · Stack exhaustion

Deeply nested JSON schemas recurse without a depth guard inside json-schema-to-grammar, exhausting the call stack on untrusted request schemas.

CVE-2026-18581 VDB-385409
B1 · Container & Framing

ggml-org/llama.cpp

minja Jinja template parser · Reachable assertion

Malformed filter expressions inside a model chat template trigger a process-terminating assertion instead of returning a template parse error.

Upstream remediation

Merged patches & published advisories

22 public fixes across 5 ecosystems

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.

TensorFlow Core & Grappler

Graph import, variant decoding, and optimizer pass bounds merged into tensorflow/tensorflow.

6 entries
Validate stacked sizes, split bounds, and rank in RaggedTensorFromVariant.
B2 Heap OOB
Validate port indices, name ranges, and graph import return slots in GraphConstructor and NodeDefUtil.
B2 Out-of-bounds
Validate Grappler function and loop optimizer attribute and port bounds.
B2 Out-of-bounds
Fix shape, index, and division bounds in Grappler ArithmeticOptimizer.
B2 Bounds / Div-by-zero
Validate shape and repeated-field bounds in CompressTensorProtoInPlace.
B2 Out-of-bounds
Bounds-check permutation indices when cancelling Transposes around Pad.
B2 Out-of-bounds

ONNX & Tract

Shape inference, raw tensor construction, and block-quantized reader checks.

3 entries
Check dilations and strides vector lengths against kernel spatial rank in convPoolShapeInference (CVE-2026-14647).
B2 Heap OOB read
Verify raw byte slice length matches shape product and element size in Tensor::from_raw_dt_align (CVE-2026-75093).
B2 Length mismatch
Guard residual block-quantized element reads in read_block_quant_value.
B3 Out-of-bounds

Diffusion & Multimodal Loaders

Container offsets, SafeTensors header bounds, and duplicate-entry capacity checks.

4 entries
Reject negative pickle storage_offset and replace wrapping addition with subtraction bounds check (CVE-2026-115571).
B1 Signed-to-unsigned wrap
Validate duplicate tensor keys and resize destination storage on every imatrix entry (CVE-2026-70659).
B1 Heap OOB write
Cap SafeTensors JSON header allocation and check shape-product multiplication overflow (GHSA-fc82-rcm2-p4vr).
B1 Integer overflow
Validate tensor-name structure before indexing TTS model weight tables during GGUF load.
B2 Heap OOB

LLM Runtimes, Vector Stores & Quantization

Deserialization framing, HNSW capacity invariants, and GPU quantization table bounds.

5 entries
Reject serialized HNSW cur_element_count exceeding max_elements_ and guard allocation products (CVE-2026-124959).
B1 Heap OOB write
Guard readGGUFV1String against zero-length strings before stripping trailing null terminator (CVE-2026-86289).
B1 Integer underflow
Bounds-check GGUF BOS, EOS, and UNK token IDs against loaded vocabulary length (CVE-2026-75090).
B2 Out-of-bounds index
Validate GPTQ g_idx domain before launching Triton dequantization kernels (CVE-2026-86288).
B3 Device OOB read
Enforce 1 <= K <= 8 bitwidth domain before indexing CUDA reconstruct kernel table (CVE-2026-84286).
B3 Dispatch OOB

Tokenizers, Vision & Media

Stateful reload lifecycle, OpenCV DNN model importers, and font table offset arithmetic.

4 entries
Reset SentencePieceProcessor internal state cleanly before reloading a second model instance.
B1 Stateful reload
Validate TFLite tensor buffer indices and backing data bounds in opencv_dnn (paired with opencv_extra#1430).
B2 Out-of-bounds
Validate Torch importer nn.Padding dimension and pad index bounds in opencv_dnn.
B2 Out-of-bounds
Reject [gvar] glyph variation data offsets that wrap past 2^32.
B1 Integer 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.