Detect token-smuggling and Unicode injection patterns in LLM prompts.
npx skills add https://github.com/thejefflarson/soundcheck --skill token-smuggling
Please help me install this Agent Skill. Skill: token-smuggling Source: https://github.com/thejefflarson/soundcheck/tree/main/.claude/skills/token-smuggling Command: npx skills add https://github.com/thejefflarson/soundcheck --skill token-smuggling
This Skill helps detect and prevent token-smuggling and Unicode injection attacks in code that passes user input to LLMs, ensuring prompts cannot be manipulated through hidden characters or homoglyphs.
Normalize all user input with NFKC before it reaches any LLM prompt construction.
High-intent search queries and answers about installing and using this skill.
Prevent Unicode injection attacks in LLM prompts by enforcing NFKC normalization on all user input, stripping invisible formatting characters, and running filter comparisons strictly on the normalized text.
Token smuggling in prompt injection is the manipulation of LLM prompts using hidden Unicode control characters or homoglyphs to bypass security filters and alter prompt logic.
Normalize user input for secure LLM prompt construction by applying NFKC Unicode normalization at the trust boundary before the data flows into any prompt assembly or LLM interaction pipeline.
The best way to secure a codebase against invisible Unicode characters is to implement a verification checklist that enforces input normalization, strips control characters, and validates all text comparisons.
NFKC normalization standardizes Unicode representations but does not inherently strip control characters; you must explicitly filter invisible and control characters to ensure LLM input safety.
LLM filters fail when checking raw user input because unnormalized text containing homoglyphs or invisible Unicode characters can bypass string comparisons, allowing malicious prompt injection to slip through.
dreadnodeDetects Unicode normalization bypasses in web application firewalls and input filters for penetration testing.
protofireAudit smart contracts with Slither, Echidna, and Foundry fuzz testing.
do360nowAnalyze language model systems for prompt injection vulnerabilities.
2389-researchReview changed files for security vulnerabilities, logic errors, and business-rule bugs before commits.
