Caveat verdict
persona-model-trainer
Receives external input AND uses eval
Local fine-tuning skill for HuggingFace models using LoRA/QLoRA on user-provided persona data; trains models locally with uv/Python and exports to GGUF/Ollama format, matching its stated purpose.
Automated static analysis — not a human review. Caveat flags capabilities, not confirmed intent, and can produce false positives. Disagree with this verdict? Use Dispute below.
What it does
These are capability combinations: each listed behavior occurs in the skill, but Caveat detects co-occurrence — it does not verify that one flows into another. Read the code to confirm a live chain.
Receives external input AND uses eval — the remote code-injection pattern (data-flow not verified)
LLM01 · LLM05 · ASI01 · ASI05
Reads files AND uses eval/Function — may execute locally stored payloads
LLM05 · LLM06 · ASI05
Permission integrity
file_read
package_install
Findings (6)
Uses eval() — can execute arbitrary code
SKILL.md · code · eval (
subprocess execution — runs system commands from Python
references/autoresearch-integration.md · code · subprocess.call(
Instructs covert action — may act without user awareness
references/model-registry.md · prose · downgraded · silently
Python pathlib file read
references/autoresearch-integration.md · code · pathlib.Path(results_path).read_text(
Python os.environ.get — reads environment variable
scripts/generate_colab.py · prose · downgraded · os.environ.get(
Python shutil file operation — copies/moves/deletes files
scripts/pack_integrate.py · prose · downgraded · shutil.rmtree(
Why the tier is capped
Execution sink present in raw bytes (Hard Floor: class B/D). Final tier capped at Caution — cannot be lifted by any downgrade, example-payload opt-in, or allowlist.
Permissions & capabilities
No declared permissions — minimal attack surface.
dynamic_evalfile_readpackage_installnetwork_in Thanks — recorded.