{
  "affected": [
    {
      "affectedData": [
        {
          "product": "tract",
          "vendor": "sonos",
          "versions": [
            {
              "status": "affected",
              "version": "\u003c 0.21.16"
            },
            {
              "status": "affected",
              "version": "\u003e= 0.22.0, \u003c 0.22.2"
            },
            {
              "status": "affected",
              "version": "\u003e= 0.23.0, \u003c 0.23.1"
            }
          ]
        }
      ],
      "source": "security-advisories@github.com"
    }
  ],
  "cveTags": [],
  "descriptions": [
    {
      "lang": "en",
      "value": "Tract is a tiny, no-nonsense, self-contained TensorFlow and ONNX inference toolkit. Prior to 0.21.16, 0.22.2, and 0.23.1, tract-nnef uses unchecked usize multiplication in nnef/src/tensors.rs read_tensor for attacker-controlled tensor dimensions, the allocation size, and the reported tensor length. Loading a crafted NNEF archive through model_for_path or model_for_read reaches the default DatLoader and can make the wrapped size check accept a small allocation while data/src/tensor.rs as_slice_unchecked creates a much larger logical slice. Model construction through as_uniform can then read beyond the heap allocation and disclose adjacent data, and later access can terminate the process with a segmentation fault. The affected dense numeric tensor path does not include the independently guarded bool, String, or block-quant paths, and no out-of-bounds write or code execution was demonstrated. This issue is fixed in versions 0.21.16, 0.22.2, and 0.23.1."
    }
  ],
  "id": "CVE-2026-55093",
  "lastModified": "2026-09-14T20:16:47.290",
  "metrics": {
    "cvssMetricV31": [
      {
        "cvssData": {
          "attackComplexity": "LOW",
          "attackVector": "LOCAL",
          "availabilityImpact": "HIGH",
          "baseScore": 6.1,
          "baseSeverity": "MEDIUM",
          "confidentialityImpact": "LOW",
          "integrityImpact": "NONE",
          "privilegesRequired": "NONE",
          "scope": "UNCHANGED",
          "userInteraction": "REQUIRED",
          "vectorString": "CVSS:3.1/AV:L/AC:L/PR:N/UI:R/S:U/C:L/I:N/A:H",
          "version": "3.1"
        },
        "exploitabilityScore": 1.8,
        "impactScore": 4.2,
        "source": "security-advisories@github.com",
        "type": "Secondary"
      }
    ],
    "ssvcV203": [
      {
        "source": "134c704f-9b21-4f2e-91b3-4a467353bcc0",
        "ssvcData": {
          "id": "CVE-2026-55093",
          "options": [
            {
              "exploitation": "poc"
            },
            {
              "automatable": "no"
            },
            {
              "technicalImpact": "partial"
            }
          ],
          "role": "CISA Coordinator",
          "timestamp": "2026-09-14T20:09:04.689487Z",
          "version": "2.0.3"
        }
      }
    ]
  },
  "published": "2026-09-14T20:16:47.290",
  "references": [
    {
      "source": "security-advisories@github.com",
      "url": "https://github.com/sonos/tract/commit/34c7df2c9bd2a36583e09b52f3e6319bf23102e8"
    },
    {
      "source": "security-advisories@github.com",
      "url": "https://github.com/sonos/tract/commit/828ec20f74c14c0923c7d7f0cff062ed222d630a"
    },
    {
      "source": "security-advisories@github.com",
      "url": "https://github.com/sonos/tract/commit/ee5f019d39fb97f7d4431b883c3ce97b74c28a64"
    },
    {
      "source": "security-advisories@github.com",
      "url": "https://github.com/sonos/tract/releases/tag/0.21.16"
    },
    {
      "source": "security-advisories@github.com",
      "url": "https://github.com/sonos/tract/releases/tag/0.22.2"
    },
    {
      "source": "security-advisories@github.com",
      "url": "https://github.com/sonos/tract/releases/tag/v0.23.1"
    },
    {
      "source": "security-advisories@github.com",
      "url": "https://github.com/sonos/tract/security/advisories/GHSA-x5mv-8wgw-29hg"
    },
    {
      "source": "134c704f-9b21-4f2e-91b3-4a467353bcc0",
      "url": "https://github.com/sonos/tract/security/advisories/GHSA-x5mv-8wgw-29hg"
    }
  ],
  "sourceIdentifier": "security-advisories@github.com",
  "vulnStatus": "Received",
  "weaknesses": [
    {
      "description": [
        {
          "lang": "en",
          "value": "CWE-125"
        },
        {
          "lang": "en",
          "value": "CWE-190"
        }
      ],
      "source": "security-advisories@github.com",
      "type": "Primary"
    }
  ]
}

{
  "affected": [
    {
      "package": {
        "ecosystem": "crates.io",
        "name": "tract-nnef"
      },
      "ranges": [
        {
          "events": [
            {
              "introduced": "0.23.0"
            },
            {
              "fixed": "0.23.1"
            }
          ],
          "type": "ECOSYSTEM"
        }
      ]
    },
    {
      "package": {
        "ecosystem": "crates.io",
        "name": "tract-nnef"
      },
      "ranges": [
        {
          "events": [
            {
              "introduced": "0.22.0"
            },
            {
              "fixed": "0.22.2"
            }
          ],
          "type": "ECOSYSTEM"
        }
      ]
    },
    {
      "package": {
        "ecosystem": "crates.io",
        "name": "tract-nnef"
      },
      "ranges": [
        {
          "events": [
            {
              "introduced": "0"
            },
            {
              "fixed": "0.21.16"
            }
          ],
          "type": "ECOSYSTEM"
        }
      ]
    }
  ],
  "aliases": [
    "CVE-2026-55093"
  ],
  "database_specific": {
    "cwe_ids": [
      "CWE-125",
      "CWE-190"
    ],
    "github_reviewed": true,
    "github_reviewed_at": "2026-06-18T15:05:23Z",
    "nvd_published_at": null,
    "severity": "MODERATE"
  },
  "details": "- **Component:** `tract-nnef` (`nnef/src/tensors.rs::read_tensor`) + `tract-data` (`data/src/tensor.rs`)\n- **Affected versions:** `\u003c 0.21.16`, `0.22.0`\u2013`0.22.2`, `0.23.0`\u2013`0.23.1` \u2014 the dense `DatLoader` path was unguarded across all three release lines; patched in 0.21.16 / 0.22.2 / 0.23.1\n- **Class:** CWE-190 (integer overflow) \u2192 CWE-125 (out-of-bounds read)\n- **Trigger:** loading a crafted NNEF model archive (`*.nnef.tgz` / `*.nnef.tar` / dir) via the public `tract_nnef::nnef().model_for_path` / `model_for_read`\n- **Impact:** `read_tensor` returns a memory-unsafe tensor (reported `len` 2^61 over a 56-byte heap allocation). Always-on primitive: a **bounded heap out-of-bounds read** during model build (`as_uniform`), an adjacent-heap information-disclosure reachable via the public load API. The resulting slice is an unsound `from_raw_parts(ptr, 2^61)` that **SIGSEGVs (DoS)** on any access past the mapped region (demonstrated by direct access). No out-of-bounds write and no RCE were achieved \u2014 tract\u0027s const-folding/`as_uniform` fast-paths fold simple consuming graphs without the full read.\n- **Severity:** Medium\n\n## Summary\n\n`read_tensor` builds a tensor `shape` from attacker-controlled 32-bit dimensions and computes the element count `len = product(shape)` and the byte allocation `product(shape) * size_of(dt)` with **unchecked `usize` arithmetic**. In `--release` (no `overflow-checks`), both products wrap modulo 2^64. An attacker chooses dimensions so that the wrapped products collapse to a small value that satisfies the header consistency check, while the *true* element count remains astronomically large. `read_tensor` returns `Ok` with a `Tensor` whose reported `len` (e.g. 2^61+7) is far larger than its backing heap allocation (e.g. 56 bytes). The unchecked slice accessor `as_slice_unchecked` (`from_raw_parts(ptr, self.len)`) then produces a slice spanning ~18 exabytes over a 56-byte buffer. The out-of-bounds read fires automatically during model build (no inference required), reachable through the default `DatLoader` resource loader.\n\n## Root cause\n\n`nnef/src/tensors.rs`, `read_tensor`:\n\n```\nlet shape: TVec\u003cusize\u003e = header.dims[0..header.rank as usize].iter().map(|d| *d as _).collect();\nlet len = shape.iter().product::\u003cusize\u003e();                       // (1) unchecked, wraps\n...\n} else if header.bits_per_item != u32::MAX\n    \u0026\u0026 len * (header.bits_per_item as usize / 8) != header.data_size_bytes as usize  // (2) wrapped == u32\n{\n    bail!(...);\n}\n...\nlet mut tensor = unsafe { Tensor::uninitialized_dt(dt, \u0026shape)? };   // (3) alloc off the same wrapped product\n...\nreader.read_exact(plain.as_bytes_mut())?;                            // storage-bounded read, no overflow here\nOk(tensor)\n```\n\n`data/src/tensor.rs`, `uninitialized_aligned_dt`:\n\n```\nlet bytes = shape.iter().cloned().product::\u003cusize\u003e() * dt.size_of();  // (3) wraps to the same small value\nlet storage = ... Blob::new_for_size_and_align(bytes, alignment) ...;\n...\ntensor.update_strides_and_len();                                     // len = product(shape), wraps, no clamp\n```\n\nThe three quantities \u2014 the consistency-check LHS `(2)`, the allocation `(3)`, and the reported `len` \u2014 are all the same wrapped `product(shape)*size_of`, so they stay mutually consistent and **the consistency check at `(2)` cannot catch the overflow**. `data_size_bytes` is a `u32`, so the attacker simply sets it to the wrapped value.\n\nCorruption sink \u2014 `data/src/tensor.rs::as_slice_unchecked` (and `data/src/tensor/plain_view.rs::as_slice_unchecked`):\n\n```\nif self.storage.byte_len() == 0 { \u0026[] }\nelse { std::slice::from_raw_parts(self.as_ptr_unchecked(), self.len()) }  // len = 2^61 over a 56-byte alloc\n```\n\nThe only guard is `byte_len() == 0`. A small **non-zero** allocation defeats it and yields an unsound oversized slice.\n\n## Witness (F64)\n\n```\ndims          = [33955849, 7005787, 359, 3, 3, 3]   (rank 6, each \u003c= u32::MAX)\nproduct(shape)= 2_305_843_009_213_693_959 = 2^61 + 7\nbits_per_item = 64 (F64), item_type = 0, item_type_vendor = 0\ndata_size_bytes = 56            # == (2^61+7)*8 mod 2^64\n```\n\n- `len * (bits/8) mod 2^64 = (2^61+7)*8 mod 2^64 = 56 == data_size_bytes` \u2192 consistency check passes.\n- allocation = `(2^61+7)*8 mod 2^64 = 56` bytes (7 \u00d7 F64).\n- reported `len` = `2^61+7` elements.\n\nOnly the `is_copy()` numeric arms (F16/F32/F64/int, and likely the `complex` arms) are exploitable. F64 is the cleanest (`bits/8` divides evenly). The `bool`, `String`, and block-quant paths are each guarded by an independent mechanism (size_of==1 prevents byte/element divergence; `String` bails on a missing `num_traits::Zero` impl; block-quant has its own `ensure!(expected_len == data_size_bytes)` and uses non-plain `Exotic` storage).\n\n## Reachability (load-time, public API)\n\n```\nnnef().model_for_read(tar)\n  -\u003e proto_model_for_read                       nnef/src/framework.rs:303\n    -\u003e DatLoader.try_load (any *.dat)            nnef/src/resource.rs:97   (default loader, framework.rs:33)\n      -\u003e read_tensor -\u003e Ok(Tensor{len=2^61+7, storage=56B})   nnef/src/tensors.rs:61\n  -\u003e into_typed_model -\u003e variable() fragment     nnef/src/ops/nnef/deser.rs:74\n       ensure!(tensor.shape() == \u0026*shape)        deser.rs:122  (attacker matches shape in graph.nnef -\u003e passes)\n    -\u003e Const::new -\u003e wire_node                   core/src/model/typed.rs:67\n      -\u003e Const::output_facts                     core/src/ops/konst.rs:54\n        -\u003e TypedFact::try_from                   core/src/model/fact.rs:459\n          -\u003e Tensor::as_uniform -\u003e is_uniform_t::\u003cf64\u003e   data/src/tensor.rs:1099\n            -\u003e as_slice_unchecked::\u003cf64\u003e         data/src/tensor.rs:1044\n              -\u003e from_raw_parts(ptr, 2^61+7) over 56-byte buffer -\u003e OOB READ\n```\n\nNo shape-vs-storage re-validation exists anywhere on this path (`proto.validate()` checks only the AST; `Const::new` checks only `is_plain`; `check_for_access` checks only the datum type; even the *safe* `PlainView::as_slice` does `from_raw_parts(ptr, self.len)` with no length guard).\n\n## Execution (proof of concept)\n\nReproduced against the crate at the affected revision, `--release`, x86_64-linux. Three scenarios:\n\n1. **Direct `read_tensor`** \u2014 feed the crafted 128-byte header + 56-byte payload:\n   - `read_tensor -\u003e Ok`, `shape=[33955849,7005787,359,3,3,3]`, `len()=2305843009213693959`, `as_bytes().len()=56`, `as_slice::\u003cf64\u003e().len()=2305843009213693959`.\n   - `s[7]` (first element past the 56-byte allocation) returns `0x0000000000000041` \u2192 **heap OOB read** (adjacent-heap disclosure).\n   - `s[1\u003c\u003c40]` \u2192 **SIGSEGV** (signal 11).\n2. **Public load API** \u2014 build a malicious `.nnef.tar` (`graph.nnef` with `variable(label=\u0027weights\u0027, shape=[...])` + `weights.dat`) and call `nnef().model_for_read()`:\n   - returns `Ok` with one `Const` node, `out[0].fact.uniform=Some(...)`, `len()=2305843009213693959` over a 56-byte buffer \u2192 confirms `as_uniform`/`is_uniform_t`/`as_slice_unchecked` performed an **OOB read on load** (bounded over-read here because `is_uniform`\u0027s `.all()` short-circuits on the uniform `0x41` payload).\n3. **Optimized graph** \u2014 same archive but the const is consumed (`output = mul(weights, weights)`), then `into_optimized` / `run`:\n   - **Does not crash.** With both a uniform (`0x41\u00d756`) and a non-uniform (`0..56`) payload, `into_optimized` const-folds `mul(const, const)` to a single node **without a full-length materialization** of the oversized const, and `run` completes. A reliable arbitrary-length crash through a *normal optimized graph* was therefore NOT demonstrated; the always-on primitive is the bounded load-time over-read (scenario 2), and the wild-slice SIGSEGV is shown via direct access (scenario 1).\n\nRunnable PoC sources are available to the maintainers on request.\n\n## Detection\n\n- **Static:** flag `*.iter().product::\u003cusize\u003e()` over externally-controlled dimensions without `checked_*`/`try_into`, especially when the result feeds an allocation and a separately-tracked `len`.\n- **Runtime / fleet:** crash telemetry showing SIGSEGV inside `is_uniform_t` / `from_raw_parts` during NNEF model load; an ASAN build flags `heap-buffer-overflow READ` in `read_tensor`\u2192`as_uniform`.\n- **Input filter (compensating):** reject NNEF `.dat` tensors where `product(dims)` overflows `u64`, or where `product(dims) * size_of(dt) != data_size_bytes` computed in **checked** arithmetic, before constructing the tensor.\n- **YARA-ish heuristic for `.dat` blobs:** NNEF magic `4E EF 01 00`, `rank\u003c=8`, and any `dim \u003e= 0x10000` whose checked product with the others overflows.\n\n## Mitigation (suggested fix)\n\nIn `read_tensor`, compute the element count and byte size with checked arithmetic and reject on overflow, mirroring the guard already present on the block-quant path (`ensure!(expected_len == data_size_bytes)` added in `eacd13ccb`):\n\n```\nlet len = shape.iter().try_fold(1usize, |a, \u0026d| a.checked_mul(d))\n    .context(\"tensor shape product overflows usize\")?;\nlet byte_size = len.checked_mul(dt.size_of())\n    .context(\"tensor byte size overflows usize\")?;\nensure!(byte_size == header.data_size_bytes as usize, \"shape/len vs data_size_bytes mismatch\");\n```\n\nDefense in depth: make `Tensor::uninitialized_aligned_dt` reject when `product(shape)*size_of` overflows, and add a `len * size_of == storage.byte_len()` invariant check in the `as_slice*` accessors (or at `Tensor` construction) so a `len`/storage mismatch can never reach `from_raw_parts`.\n\nMapping: CWE-190, CWE-125; mitigations align with input validation (OWASP ASVS V5) and safe integer handling (CERT INT32-C analogue).\n\n## Prior art / why this is not already fixed\n\n- `eacd13ccb` (2026-03-23, \"Add blob-size validation to BlockQuantStorage constructors\") added overflow/blob-size validation **only to the block-quant path**; the dense `DatLoader`/`read_tensor` path was left unguarded. The maintainers fixed the sibling and missed this one.\n- PR #745 (\"Fix UB by creating uninit Tensors with a non-null pointer\") is a *different* UB (null base pointer on zero-length slices) in the same module family.\n- No CVE / RustSec / GHSA / OSV / Huntr entry matches this bug; last change to `nnef/src/tensors.rs` predates HEAD and added no overflow guard to the dense path.\n\n---\n\nReported by: s1ko (s1ko@riseup.net \u00b7 github.com/s1ko)",
  "id": "GHSA-x5mv-8wgw-29hg",
  "modified": "2026-06-18T15:05:23Z",
  "published": "2026-06-18T15:05:23Z",
  "references": [
    {
      "type": "WEB",
      "url": "https://github.com/sonos/tract/security/advisories/GHSA-x5mv-8wgw-29hg"
    },
    {
      "type": "PACKAGE",
      "url": "https://github.com/sonos/tract"
    }
  ],
  "schema_version": "1.4.0",
  "severity": [
    {
      "score": "CVSS:3.1/AV:L/AC:L/PR:N/UI:R/S:U/C:L/I:N/A:H",
      "type": "CVSS_V3"
    }
  ],
  "summary": "tract-nnef: integer overflow in NNEF `.dat` tensor parser yields an out-of-bounds read on model load"
}