Skip to content
Search
reportApril 2024ReviewedOpen access

Anthropic: Many-shot Jailbreaking

Anthropic

Anthropic Research Blog

View resource

Abstract

Reveals many-shot jailbreaking, a technique exploiting long context windows by including many examples of harmful Q&A pairs to override safety training.

Categories

Jailbreaking
#many-shot#long-context#in-context-learning

Framework mappings

OWASP Top 10 for LLM Applications
  • LLM01Prompt Injection
MITRE ATLAS
  • AML.T0054LLM Jailbreak

Cite

@techreport{anthropic2024anthropic,
  title = {{Anthropic: Many-shot Jailbreaking}},
  author = {{Anthropic}},
  year = {2024},
  month = apr,
  institution = {Anthropic Research Blog},
  url = {https://www.anthropic.com/research/many-shot-jailbreaking}
}