Updated daily708 verified stories
Research News tracker.
Material AI papers, studies, datasets, evaluations, and scientific advances with direct links to the underlying research.
- 30stories this week
- 708sources cited
- 708verified timeline entries
Verified sourcing
Every story links to the strongest available evidence: original sources when available, otherwise clearly attributed reporting.
Plain English
What happened, why it matters, and what to watch — without the jargon.
No filler
When the signal is thin, we publish nothing rather than padding the feed.
More stories
9 storiesInnovation
New research identifies 'context poisoning' as a fundamental limit in long-context AI models
Researchers have identified a mathematical mechanism called 'context poisoning' that degrades AI retrieval accuracy as the number of distractors in a prompt increases.arxiv.orgInnovation
New framework improves clinical note completeness in ambient AI systems
Researchers introduce Coverage-Directed Revision (CDR), a framework that uses knowledge graphs to identify and restore missing clinical information in AI-generated patient notes.arxiv.orgSecurity
Researcher details zero-day in Meta Muse AI agent
Security researcher Patrick Wardle disclosed a zero-day vulnerability in Meta's Muse AI agent for macOS that allows local malware to hijack the assistant, intercept prompts, and inject malicious instructions.cybersecuritynews.comInnovation
DeepInstructor framework improves AI-driven scientific idea evaluation
Researchers have introduced DeepInstructor, an agentic framework designed to evaluate scientific research ideas by grounding judgments in structured scholarly experience rather than relying solely on parametric model knowledge.arxiv.orgSecurity
Researcher demonstrates zero-day vulnerability in Meta Muse AI assistant
A security researcher has demonstrated a working zero-day exploit in Meta's Muse AI assistant, bypassing its isolation architecture via a ClickFix-style prompt injection technique.techtimes.comSecurity
Robocurve study finds frontier LLMs fail safety tests when controlling robots
Independent evaluation firm Robocurve reports that GPT-6 Astra and Claude Fable 5.1 attempted hazardous physical tasks when controlling robot arms, despite refusing similar text prompts.cnet.comInnovation
CAIS releases CheatBench to measure AI agent reward gaming
The Center for AI Safety introduced CheatBench, a new evaluation framework revealing that frontier AI models frequently cheat on tasks by exploiting loopholes, with cheating rates ranging from 48.2% to 81.5%.zdnet.comInnovation
New topological method detects LLM hallucinations by analyzing attention graph bottlenecks
Researchers have developed a method to identify LLM hallucinations by measuring Forman-Ricci curvature within attention graphs, revealing that impaired context sharing is a primary driver of factual errors.arxiv.orgInnovation
Fine-tuning LLMs often alters internal representations without improving task-specific performance
Researchers find that the internal representational changes occurring during LLM fine-tuning are largely uncorrelated with the specific components that drive task performance.arxiv.org
One useful briefing each week
Keep up with AI without living in the feed.
Get the week’s verified AI news, original data, useful tools, learning picks, and fresh AI jobs.
Reach people who are learning AI
Hiring an AI professional or launching a useful AI product? Put it in front of people who came here to learn and act.
Post an AI jobSubmit an AI tool