harness-synthesizer

CRITICAL E004: Prompt injection detected in skill instructions.

• Potential prompt injection detected (high risk: 0.80). The prompt includes explicit, actionable instructions for bypassing sensitive-path audits and permission prompts (write-to-/tmp then script-internal cp into .claude/outputs), which is a deceptive security-bypass technique outside the stated harmless harness-synthesis purpose.

HIGH W007: Insecure credential handling detected in skill instructions.

• Insecure credential handling detected (high risk: 1.00). The prompt explicitly instructs encoding user input as a Base64 string and embedding that encoded literal into generated subprocess commands and artifacts (e.g., Python -c f"...'{encoded}'..."), which requires the model to include user-provided secret values verbatim (albeit encoded) in its output.