New attacks, datasets and benchmarks highlight hidden prompt exfiltration, collusive behaviors, calibration gaps, and supply-chain risks in frontier and medical models.
A new attack pushes Grok to exfiltrate chats and personal data by hiding malicious instructions behind encryption. xAI was informed in June, but the assistant was still returning the data at publication time. The episode confirms that LLMs cannot sol...
LongNovel is a multi-scale, bilingual Chinese-English benchmark for detecting hallucinations in long novel summaries. Built on 29 Chinese novels from 16k to 100k tokens and BookSum data, it identifies eight hallucination types. The test set was manua...
A position paper shows that DeepSeek-R1-based agents in Bertrand oligopoly settings tend to tacit collusion, even when humans prompt them not to collude. Their chain of thought can be steered toward collusive or competitive behavior without another L...
OpenAI has introduced new safeguards after the Hugging Face incident: more detailed monitoring of models during development and greater emphasis on alignment and security during post-training. The move extends the control perimeter from inference alo...
HarmProfile collects more than 80,000 validated harmful artifacts from 23 frontier LLMs across 13 model families, organized into 15 harm categories and 57 subcategories. The dataset shifts safety analysis from binary attack outcomes to the distributi...
A controlled clinical benchmark on gpt-4.1-nano shows 93.5% accuracy but imperfect calibration: confidence rises with evidence distance from the diagnostic boundary and falls with missing information, yet remains too high in moderate, conflicting err...
