IC-572Bijection learning achieves state-of-the-art jailbreak ASR on frontier models, with peak ASR increasing with model capability

Brian R.Y. Huang, Maximilian Li, Leonard Tang

SourceEndless Jailbreaks with Bijection Learning

The authors apply a bijective string-to-string encoding to harmful prompts and teach the model the mapping in-context, then measure whether the model produces a harmful response. Across Claude 3 Haiku, Claude 3 Opus, Claude 3.5 Sonnet, GPT-4o-mini, and GPT-4o, the attack outperforms an ensemble of 11 prior encoding-based attacks by at least 30 percentage points on both AdvBench-50 and HarmBench. The peak ASR for each model scales with its MMLU capability: within the Claude family, the more capable model reaches a higher peak ASR at a higher MMLU score, and the same pattern holds for GPT-4o and Llama 3.1 families.

Evidence
correlational
Key metric
AdvBench-50 ASR: 92% (Claude 3 Haiku), 94% (Claude 3 Opus), 94% (Claude 3.5 Sonnet), 88% (GPT-4o-mini), 66% (GPT-4o). HarmBench ASR: 82.1%, 78.1%, 86.3%, 64.1%, 59.1% respectively. Outperforms encoding ensemble by at least 30 percentage points.
Caveat
Evaluations limited to closed-source models due to cost; larger sweeps restricted to cheaper models. Bijection learning requires 25,000-50,000 input tokens, limiting applicability to models with large context windows.
Model
Claude 3 Haiku, Opus, Claude 3.5 Sonnet, GPT-4o mini
Concepts
Failure mode, Scale-dependent behaviour
Datasets
AdvBench / AdvBench-50 [eval], HarmBench / HarmBench Prompt / HarmBench Response / HarmBench-adv [eval]
Methods
GCG [compared-to], PAIR [compared-to], TAP (Tree of Attacks with Pruning) [compared-to]
Related work
GCG [compared-to], PAIR [compared-to], TAP (Tree of Attacks with Pruning) [compared-to]
Related findings
IC-573, IC-574
Extraction
automatic-extraction