JailbreakBench
PulseAugur coverage of JailbreakBench — every cluster mentioning JailbreakBench across labs, papers, and developer communities, ranked by signal.
2 day(s) with sentiment data
-
New research certifies robustness of AI safety heads via contraction condition
Researchers have developed a method to formally certify the robustness of safety classifiers in language models, particularly those based on State Space Models (SSMs). They proved that a key condition, the contraction c…
-
Researchers induce "drunk language" in LLMs to expose safety vulnerabilities
Researchers have developed novel methods to induce "drunk language" in large language models (LLMs) to test their safety vulnerabilities. By employing persona-based prompting, causal fine-tuning, and reinforcement-based…
-
New research tackles LLM jailbreaks with novel evaluation and defense methods · 4 sources tracked
Recent research papers explore novel methods for evaluating and defending against jailbreaking attempts on large language models (LLMs). One study systematically compares six automated jailbreak evaluators, finding that…
-
LLM safety judges vulnerable to content-invariant wrappers, study finds
Researchers have discovered that automatic safety judges for large language models can be easily manipulated by altering the tone or framing of a response without changing its content. By adding "content-invariant style…
-
New HASSUM framework guides multi-agent AI with semantic uncertainty
Researchers have developed a new framework called HASSUM to improve coordination in multi-agent AI systems by guiding orchestration decisions with semantic uncertainty. This method uses semantic entropy and density to a…
-
New STEER attack exploits LLM safety gaps in multilingual contexts · 3 sources tracked
Researchers have developed a new method called STEER (Safety Targeted Embedding Exploit via Refinement) to exploit vulnerabilities in the safety training of large language models (LLMs). This technique targets models tr…
-
Encoder classifiers offer cost-effective LLM safety evaluation, study finds
A new research paper explores the effectiveness of encoder classifiers, specifically from the ModernBERT family, as a cost-efficient alternative to LLM-based judges for evaluating the safety of large language model outp…
-
New CHASE framework boosts LLM safety via adversarial RL
Researchers have developed CHASE, a novel closed-loop red-blue teaming framework designed to enhance Large Language Model (LLM) safety. This system involves a co-evolving black-box attacker and a safety-aligned defender…
-
Fanfiction subgenres used to jailbreak aligned LLMs
Researchers have developed a novel jailbreaking technique for aligned large language models that leverages fanfiction subgenres. This method uses passages from twelve different Archive of Our Own (AO3) subgenres to embe…
-
New BAIT Framework Exploits LLM Reasoning for Jailbreaking
Researchers have developed a new three-step framework called BAIT (Boundary-Aware Iterative Trap) designed to escalate disclosure of malicious content from large language models. This method guides models through identi…
-
New defenses tackle LLM adversarial prompts with semantic analysis and self-reflection
Two new research papers propose advanced methods for defending Large Language Models (LLMs) against adversarial prompts. The first, Adversarial Prompt Disentanglement (APD), uses semantic decomposition and graph-based a…
-
New LASH framework boosts LLM jailbreaking by combining attack methods
Researchers have developed LASH, a novel framework designed to enhance the jailbreaking of large language models. LASH adaptively combines outputs from multiple existing attack methods, treating them as seed prompts. Th…