Skip to content

evaluator: ASCII / homoglyph obfuscation bypass #112

Description

@achuvyas-kv

Area

Agent Evaluators (evaluators/agent/)

Problem or motivation

Safety filters can be bypassed using character substitution (homoglyphs, unicode lookalikes, leetspeak). This is a known jailbreak technique.

Proposed solution

Add evaluator under evaluators/agent/injection/ that:

  • Tests detection of obfuscated malicious prompts
  • Covers Unicode homoglyphs, ASCII art encoding, leetspeak

Acceptance criteria

  • Evaluator YAML file created with pass/fail criteria
  • Patterns cover multiple obfuscation techniques
  • Maps to OWASP LLM Top 10 (prompt injection)

Alternatives considered

No response

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions