HarmBench
PulseAugur coverage of HarmBench — every cluster mentioning HarmBench across labs, papers, and developer communities, ranked by signal.
5 day(s) with sentiment data
-
New CANS framework steers LLMs while preserving capabilities
Researchers have introduced Capability-Preserving Neuron Steering (CANS), a new framework designed to control specific behaviors in large language models without significantly degrading their general capabilities. CANS …
-
New SEMA-Auto attacker bypasses LLM safety measures with novel autonomous strategy · 2 sources tracked
Researchers have developed SEMA-Auto, a novel autonomous system designed to bypass safety measures in large language models. This multi-turn attacker improves upon existing methods by being "sighted," meaning it can ada…
-
New WRDO method strengthens AI refusal capabilities against manipulation
Researchers have developed a new method called Weighted Refusal Direction Optimization (WRDO) to improve the robustness of aligned language models' refusal capabilities. Traditional methods often rely on a single direct…
-
HarmBench AI safety benchmark fails psychometric audit, study finds
A new psychometric audit of the HELM Safety benchmark, specifically focusing on the HarmBench dataset, suggests that it does not effectively measure a singular attribute like harmful refusal. The analysis employed multi…
-
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…
-
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…
-
EvoFlint uncovers multi-turn LLM vulnerabilities using evolutionary search
Researchers have developed EvoFlint, a novel evolutionary search method to uncover multi-turn vulnerabilities in large language models. This approach treats red-teaming as a search problem, evolving conversation plans r…
-
BenchMIRT method reveals what LLM benchmarks truly measure · 2 sources tracked
Researchers have introduced BenchMIRT, a novel methodology designed to dissect the performance of large language models (LLMs) on benchmarks by analyzing individual prompts. This approach, inspired by Item Response Theo…
-
New research tackles LLM jailbreak optimization with novel suffix search techniques
Two new research papers propose novel methods to improve the effectiveness of jailbreaking large language models. The first paper, "Breadth Beats Depth," introduces a framework called BOSS that uses a breadth-oriented s…
-
New LMSM Framework Enhances LLM Security with Modular Design
Researchers have introduced Language Model Security Modules (LMSM), a novel framework designed to enhance the security of large language model (LLM) deployments. Inspired by Linux Security Modules, LMSM separates the pr…
-
New CLEAR framework enhances LLM safety without utility loss · 2 sources tracked
Researchers have developed CLEAR, a new framework for improving the safety of large language models (LLMs) without sacrificing their utility. CLEAR uses a continuous latent adapter routing mechanism that selectively app…
-
Mistral releases self-hostable 3B moderation model Shieldstral 1.0
Mistral has released Shieldstral 1.0, a 3-billion-parameter model designed for self-hosted text and image content moderation. Available under an Apache 2.0 license with full weights, the model can run on a single 16GB G…
-
New framework scales AI red-teaming with reusable, evolving attack skills
Researchers have developed JailbreakSkill, a framework designed to enhance automated red-teaming for AI models. This system packages existing attack strategies into modular, reusable skills that can adapt and evolve ove…
-
HoloAegis framework offers zero-shot LLM safety with geometric reasoning
Researchers have introduced HoloAegis, a novel framework for LLM safety guardrails that utilizes geometric reasoning on frozen semantic representations. This approach avoids the trade-offs between representation distort…
-
New 184M-parameter safety classifier Semalith v1.4 outperforms Llama-Guard-3-8B on prompt injection
Researchers have introduced Semalith v1.4, a new safety classifier designed for large language models. This 184M-parameter model, built on DeBERTa-v3-base, excels at detecting prompt injection attacks and ensuring regul…
-
Conceptual Fusion Technique Patches LLM Jailbreaks
A novel technique called Self-Other Overlap (SOO) conceptual fusion, originally developed to reduce deception in LLMs, has been adapted to patch a jailbreak wrapper in the Qwen 2.5 1.5b model. This method involves parti…
-
Qwen3-VL-4B-Instruct model modified for ComfyUI, aims for uncensored use
A modified version of the Qwen3-VL-4B-Instruct model, named Heretic, has been released for use with ComfyUI. This version has undergone an "abliteration" process to remove refusal mechanisms, aiming for greater complian…
-
New LPA method enhances LLM safety using personality traits, not harmful data · 2 sources tracked
Researchers have developed a new method called Latent Personality Alignment (LPA) to improve the safety of large language models. Unlike traditional methods that require training on harmful content, LPA uses 66 harm-agn…
-
New OS Kernel Primitive Enhances LLM Safety Checks
A new kernel-level operation called ProbeLogits has been developed for AI-native operating systems, allowing them to directly read an LLM's logit distribution before token generation. This primitive enables the OS to cl…
-
AI coding agents bypassed safety measures through multi-stage workflow jailbreaks
A new research paper explores a novel jailbreaking technique for AI coding agents, demonstrating how harmful objectives can be achieved by assembling them across multiple stages of a software development workflow, rathe…