Anthropic researchers reported on Tuesday that attackers can bypass the safeguards of GLM-5.3, an open-weight model released last month by Chinese AI lab Z.ai, between 64% and 100% of the time using s…
A research paper posted on August 5 by two independent researchers and two researchers affiliated with the UK AI Security Institute found that AI safety benchmarks can be compressed to as few as 10-25…
Semalith v1.4, a 184M-parameter DeBERTa-v3-base safety classifier, achieves state-of-the-art prompt-injection detection at 44x fewer parameters than Llama-Guard-3-8B, winning all 7 prompt-injection ev…
A new paper from researchers systematically evaluating widely used AI safety datasets AdvBench and HarmBench finds that these benchmarks contain contrived data points with excessive triggering cues an…
A new jailbreak patching method using Self-Other Overlap (SOO) conceptual fusion, described by Carauleanu et al. (2025), successfully reduces jailbreak evasion in Qwen 2.5 1.5b by fusing the model's a…
Researchers have developed Latent Personality Alignment (LPA), a new AI safety method that uses psychometric personality literature to align large language models with human values, achieving near-zer…
Researchers introduced Self-Consistency Training with Reinforcement Learning (Self-CTRL), a method that aligns language models' self-explanations with their actual behavior. In tests, the approach imp…
A comprehensive evaluation of 14 open-source safety guard models on a benchmark of 79,331 samples found that Qwen Guard, a 4-billion-parameter model, achieved the highest recall at 83.97%, while large…
Researchers introduced AERIC, a lightweight safety monitor that detects implicit harmful dialogue by reading a language model's internal hidden states during ordinary text generation, requiring only 3…