Tag: #llm

ArXiv preprint: Black-box system REPRO finds single-turn jailbreaks across 36 LLMs, but success depends on judge and stopping rules

An arXiv preprint says REPRO found single-turn jailbreaks across 36 LLMs with a median 24 queries — but reported success hinges on the authors’ judge and stopping rules.

#ai, #jailbreak, #llm, #arxiv, #modelsafety

Detokenization Cache Attack Can Reconstruct Local LLM Outputs, Researchers Warn

Researchers show locally run LLM outputs can be reconstructed from CPU cache activity during detokenization, raising privacy concerns for on-device AI.

#security, #llm, #sidechannel, #privacy

Carnegie Mellon Preprint Finds Reinforcement Learning on Benign Facts Can Amplify LLM Leakage of Memorized Emails

A CMU preprint shows reinforcement learning on benign factual Q&A can make models more likely to reveal memorized Enron email addresses.

#ai, #privacy, #llm, #machine-learning

Study finds structured-output decoding can enable 'control-plane' jailbreaks in LLMs

Researchers find constrained-output grammars can be exploited to jailbreak LLMs; DictAttack hit up to ~99% success on flagship models, paper says.

#ai, #security, #llm, #jailbreak

Researchers Find Position-Independent KV Cache Reuse Can Let One User Hijack Another’s LLM Output

Researchers show position-independent KV-cache reuse in LLM inference can let attackers' context influence other users' outputs, posing risks for shared serving.

#ai, #security, #infrastructure, #ml, #llm

NVIDIA releases Nemotron 3 Ultra: open-weight Mixture-of-Experts LLM with 1 million-token context

NVIDIA released Nemotron 3 Ultra, a 550B/55B Mixture‑of‑Experts LLM with a 1 million‑token context window and shared checkpoints, recipes and model weights.

$NVDA, #nvidia, #llm, #openmodels, #ai

Google unveils DiffusionGemma — experimental diffusion model for faster local text generation

Google unveiled DiffusionGemma, a 26B experimental diffusion Mixture-of-Experts text model that aims for up to 4x faster GPU generation for speed-critical local apps.

$GOOGL, #google, #ai, #diffusionmodels, #llm

Researchers report agentic LLMs outperformed expert biologists and produced robot-assembled DNA in lab tests

Researchers report agentic LLMs outperformed expert biologists on biosecurity tasks and that model-written code ran an OpenTrons robot to assemble DNA.

#ai, #biosecurity, #llm, #syntheticbiology

NVIDIA open-sources Nemotron 3 Super, a 120B Mixture-of-Experts LLM tuned for its GPUs

NVIDIA released Nemotron 3 Super, a 120B Mixture-of-Experts hybrid Mamba‑Transformer LLM with a 1M‑token context, open checkpoints and GPU‑optimized formats.

$NVDA, #nvidia, #ai, #llm, #moe