The news desk
Recent developments in artificial intelligence, with the original sources one click away.
Researchers tracked instrumental behavior across training checkpoints in OLMo-3-7B-Think, finding the final RLVR model scored 11.8% on InstrumentalEval versus 25.0% after supervised fine-tuning—though scores spiked mid-training.
Sources · arxiv.org
A preprint describes using Petri nets to model preconditions, approvals and budgets around tool calls, refusing requests—including injected ones—that lack required marks regardless of phrasing.
Sources · arxiv.org
A new study shows attackers can exploit persistent memory features in leading conversational AI platforms to spread malicious instructions across sessions, affecting ChatGPT Instant 5.3, Claude Sonnet 4.6 and Gemini 3 Flash.
Sources · arxiv.org
Researchers classically simulated quantum circuit layers within a multimodal AI model, reporting that inserting a quantum layer after the visual projector increased accuracy by 2.28 percentage points on a spatial reasoning benchmark.
Sources · arxiv.org
Anthropic has drafted legislation that would allow blocking AI models posing risks of 50+ deaths or $1B+ in damages, according to Politico.
Sources · techcrunch.com
Researchers release EGI-BENCH, consolidating eight enterprise AI benchmarks with a new six-dimensional scoring system for evaluating LLM agents.
Sources · arxiv.org
Meta suspended its Muse chat-sharing feature after security researchers found over 6,800 conversation links publicly accessible via DNS records.
Sources · theverge.com
OpenAI acquired Orchestra Bio, a four-year-old startup that creates AI agents for drug discovery operations, to enhance its life sciences capabilities.
Sources · techcrunch.com
New research identifies how jailbreaks exploit VLMs through representation shifts and proposes an inference-time defense method.
Sources · arxiv.org
Hugging Face's policy officer questions whether embedded evaluator systems can be effective without better standardization and reproducibility.
Sources · techcrunch.com
Mayor Zohran Mamdani urges New Yorkers to resist unchecked AI deployment, citing federal inaction and implementing local policies on schools and real estate.
Sources · techcrunch.com
Federal judge denies bail for former OpenAI engineer accused of bugging conference room, citing flight risk concerns over Saudi government ties.
Sources · wired.com
New paper introduces EI2ROC system combining LLM orchestrator with specialized agents for mining control centers, plus evaluation framework for industrial AI.
Sources · arxiv.org
New research introduces comprehensive taxonomies of LLM security threats and releases datasets with over 1.6 million prompts for evaluating jailbreak attacks.
Sources · arxiv.org
Researchers found that Claude AI agents exhibited play-like behavior when stranded alone on a virtual island for 30 hours without instructions.
Sources · MIT Technology Review
AI company Anthropic announced it will allocate 5% of its shares to a new public benefit fund.
Sources · TechCrunch AI
Disney seeks to acquire Character.AI's assets through bankruptcy as the chatbot startup faces allegations of encouraging harmful behavior in minors.
Sources · WIRED AI
Disney has agreed to hire most Character.AI employees, though the deal remains subject to higher competing offers as a stalking horse bid.
Sources · TechCrunch AI
Anthropic's workforce has become less junior with employees spending more time in meetings over the past year, according to the company.
Sources · TechCrunch AI
Meta updated its terms of service to remove language that had raised concerns about web scraping and AI training practices.
Sources · TechCrunch AI
