Agentic AI
New Framework Automates Jailbreak Testing, Keeping LLM Safety Benchmarks Current
A multi-agent system translates attack research into standardized, reproducible code to help security teams evaluate AI robustness at scale.
Jailbreak methods for large language models are outpacing the ability to benchmark them fairly. Researchers have introduced Jailbreak Foundry (JBF), a system designed to keep security assessments synchronized with the pace of attack innovation.
The Core Problem#
LLM attacks evolve rapidly, but existing benchmarks struggle to keep up. Papers describing new jailbreak techniques typically use different test datasets, evaluation protocols, and judging criteria, making it nearly impossible to compare attack effectiveness across studies. This fragmentation leaves security teams with outdated robustness estimates and inconsistent measurements.
How Jailbreak Foundry Works#
JBF addresses this through three components:
- JBF-LIB: Provides shared code contracts and reusable utilities that reduce duplication
- JBF-FORGE: A multi-agent workflow that automatically translates academic papers into executable attack modules
- JBF-EVAL: Standardizes how attacks are measured across all implementations
By automating the conversion from research paper to runnable code, the system cuts custom implementation overhead by more than 50% and achieves an 82.5% code reuse rate across attacks.
Results and Scale#
The team reproduced 30 different jailbreak attacks and achieved high fidelity—the gap between reproduced and originally reported attack success rates averaged just 0.26 percentage points. Using a single GPT-4o judge, they then evaluated all 30 attacks consistently across 10 victim models on the AdvBench standard, something impossible to do fairly when attacks are scattered across incompatible codebases.
Why It Matters#
By creating a shared infrastructure for attack evaluation, JBF enables a "living benchmark" that stays current as new threats emerge. Rather than waiting months for benchmark papers to incorporate the latest research, security teams can immediately test new attack papers within the unified framework, keeping defenses aligned with the actual threat landscape.
