local-finetune

HIGH W007: Insecure credential handling detected in skill instructions.

• Insecure credential handling detected (high risk: 1.00). The pipeline ingests raw chat history (~/.claude/history.jsonl) and writes message content verbatim into the DB and exported JSONL training files without redaction, which will embed any API keys/tokens/passwords present in history into outputs and fine-tuning data.

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.85). Runtime LLM context is fed from outsider-authored free text in claude_history via read_json('~/.claude/history.jsonl', ...) where display becomes content, and then training_candidates exports instruction/output into JSONL messages consumed by mlx_lm.lora.