Boundary Point Jailbreaking of Black-Box LLMs

TL;DR

Boundary Point Jailbreaking (BPJ) is a fully automated black-box attack method that effectively bypasses advanced safeguards in large language models by using minimal feedback and boundary detection techniques, challenging current defense strategies.

Contribution

BPJ introduces a novel automated attack approach that operates with minimal information, successfully evading industry-deployed safeguards against large language models.

Findings

BPJ can develop universal jailbreaks against Constitutional Classifiers.

BPJ successfully attacks GPT-5's input classifier without human seeds.

BPJ is difficult to defend against in individual interactions.

Abstract

Frontier LLMs are safeguarded against attempts to extract harmful information via adversarial prompts known as "jailbreaks". Recently, defenders have developed classifier-based systems that have survived thousands of hours of human red teaming. We introduce Boundary Point Jailbreaking (BPJ), a new class of automated jailbreak attacks that evade the strongest industry-deployed safeguards. Unlike previous attacks that rely on white/grey-box assumptions (such as classifier scores or gradients) or libraries of existing jailbreaks, BPJ is fully black-box and uses only a single bit of information per query: whether or not the classifier flags the interaction. To achieve this, BPJ addresses the core difficulty in optimising attacks against robust real-world defences: evaluating whether a proposed modification to an attack is an improvement. Instead of directly trying to learn an attack for a target harmful string, BPJ converts the string into a curriculum of intermediate attack targets and then actively selects evaluation points that best detect small changes in attack strength ("boundary points"). We believe BPJ is the first fully automated attack algorithm that succeeds in developing universal jailbreaks against Constitutional Classifiers, as well as the first automated attack algorithm that succeeds against GPT-5's input classifier without relying on human attack seeds. BPJ is difficult to defend against in individual interactions but incurs many flags during optimisation, suggesting that effective defence requires supplementing single-interaction methods with batch-level monitoring.