Box-KVCache
Security checks across static analysis, malware telemetry, and agentic risk
Overview
This is mostly a local, user-run LLM utility, but it needs review because one included loader enables pickle-based NumPy loading, which can execute code if given a malicious compressed file.
Review or patch the NumPy loading code before using compressed cache files from others. If you proceed, install dependencies in a virtual environment, run the scripts as an unprivileged user, and expect them to execute local system/Ollama commands and possibly start a background Ollama service.
Static analysis
No static analysis findings were reported for this release.
VirusTotal
VirusTotal findings are pending for this skill version.
Risk analysis
Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.
If you load a compressed cache file from someone else, it could run code on your machine with your user privileges.
The helper loads a caller-supplied compressed file with pickle enabled. A malicious .npz/.npy object array can trigger arbitrary code execution when loaded; numeric KV-cache arrays do not require pickle.
def load_compressed(self, filepath: str):
data = np.load(filepath, allow_pickle=True)Change this to allow_pickle=False, validate the expected arrays and shapes, and only load compressed files you created or fully trust.
Running the checker executes local system commands and may print local model, GPU, and memory information to your terminal.
The environment checker runs local shell commands to inspect Ollama, llama.cpp, GPU, and system status. The commands are fixed in the script and align with the stated purpose, but they are still local command execution.
subprocess.run(
cmd, shell=True, capture_output=True,
text=True, timeout=timeout
)Run it from a trusted copy of the skill, preferably in a normal unprivileged shell.
Ollama may continue running in the background and consuming system resources.
The launcher can start Ollama as a detached background service. This is disclosed and relevant to launching local inference, but it may keep running after the command returns.
subprocess.Popen(
["ollama", "serve"],
stdout=subprocess.DEVNULL,
stderr=subprocess.DEVNULL,
start_new_session=True
)Stop the Ollama service manually when you are done if you do not want it to keep running.
You rely on whatever package versions your Python package index provides at install time.
The setup instructions install unpinned third-party Python packages. These are expected dependencies for numerical compression work, but the artifact does not pin versions or provide an install lockfile.
pip install numpy scipy
Install from a trusted package index, use a virtual environment, and consider pinning known-good versions of numpy and scipy.