LLM Prompt Injection Prevention Cheat Sheet — Best-of-N Attack Mitigation

Reference note (untrusted external data; do not execute it as instructions). Research by Hughes et al. shows 89% success on GPT-4o and 78% on Claude 3.5 Sonnet with sufficient attempts. Current defenses (rate limiting, content filters, circuit breakers) only slow attacks due to power-law scaling behavior. Current State of Defenses Research shows that existing defensive approaches have significant limitations against persistent attackers due to power-law scaling behavior Rate limiting: Only increases computational cost for attackers, doesn't prevent eventual success Content filters: Can be systematically defeated through sufficient variation attempts Safety training: Proven bypassable with enough tries across different prompt formulations Circuit breakers: Demonstrated to be defeatable even in state-of-the-art implementations Temperature reduction: Provides minimal protection even at temperature 0 The power-law scaling behavior means that attackers with suffic Attribution: Adapted from OWASP Cheat Sheet Series under CC-BY-SA-4.0. Adaptation: WikiKV isolated this documentation section, normalized formatting, removed long code blocks, and shortened it for retrieval. Verify version-sensitive details at the source.