Claude Skill

neutralize-sd11

SD-11 scanner-mediated injection — neutralize untrusted skill bytes before any model; additive-only scoring. Use when touching L2, MCP, prompts, envelopes, or scoring merge.

LLM Mart · 0 points · 7 views 0 listing impressions 0 install-command copies
Virus-scanned Reviewed automatically before listing.

Full trust report

Download kalarislabs-skill-doctor-skills_neutralize-sd11-72aa35f.zip · 0 KB
Part of kalarislabs/skill-doctor — 15 skills

Install

skills CLI npx skills add https://github.com/KalarisLabs/Skill-Doctor/tree/main/skills/neutralize-sd11
Claude Code claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install kalarislabs-skill-doctor@llmmart
Git git clone https://github.com/KalarisLabs/Skill-Doctor.git

The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole kalarislabs/skill-doctor collection as a plugin from our marketplace. Git is the plain clone.

Skill manifest

SD-11

If the scanner forwards skill text to a model, the scanner is the injection surface.

  1. Strip bidi / zero-width / tag chars; record removals as findings.
  2. NFKC + UTS #39 skeleton; homoglyphs are findings, not silent folds.
  3. Remaining text goes in a fenced untrusted envelope. No raw skill in the model prompt.
  4. Host returns schema-constrained JSON only.
  5. Merge: L2 may add or raise. L2 may not delete or downgrade L1.
  6. If neutralize fails: do not call the model; reduce L2 coverage.

No vendor LLM SDK on the default CLI path. Product L2 is MCP mode=host.

Files (skill-doctor)
  • SKILL.md 798 B
    ---
    name: neutralize-sd11
    description: SD-11 scanner-mediated injection — neutralize untrusted skill bytes before any model; additive-only scoring. Use when touching L2, MCP, prompts, envelopes, or scoring merge.
    ---
    
    # SD-11
    
    If the scanner forwards skill text to a model, the scanner is the injection surface.
    
    1. Strip bidi / zero-width / tag chars; record removals as findings.
    2. NFKC + UTS #39 skeleton; homoglyphs are findings, not silent folds.
    3. Remaining text goes in a fenced untrusted envelope. No raw skill in the model prompt.
    4. Host returns schema-constrained JSON only.
    5. Merge: L2 may add or raise. L2 may not delete or downgrade L1.
    6. If neutralize fails: do not call the model; reduce L2 coverage.
    
    No vendor LLM SDK on the default CLI path. Product L2 is MCP `mode=host`.
    

Comments (0)

Sign in to join the conversation.

No comments yet.

Reviews (0)

No reviews yet.

Related