OpenAI's new reasoning technique worries AI safety experts
OpenAI's new Astra model uses an 'opaque recurrence' reasoning technique that AI safety researchers say could make its chain of thought much harder to monitor.
Sections
AI & Models
528 storiesOpenAI's Astra model is on the way — and very good at hacking
OpenAI said its upcoming Astra model is the first to meet its "critical cybersecurity threshold," scoring a perfect score on ExploitBench and finding two zero-day vulnerabilities in an internal test.
Anthropic's new Fable release is cheaper, less restrictive
Anthropic released Fable and Mythos 5.1 on Tuesday, twinned model updates that cut token costs and reduce false-positive safeguard restrictions alongside performance gains.
Pentagon launches custom ChatGPT and Grok for personnel
The Pentagon has launched government versions of ChatGPT and Grok, giving up to 3 million personnel access to AI tools tailored to "warfighter needs."
Anthropic paper offers an early look at AI that improves its own alignment training
An Anthropic fellow's new paper shows automated AI systems reliably improving a model's performance on 10 alignment benchmarks, offering an early glimpse of self-improving AI research.
Open-weight AI companies become the Valley's hottest acquisition targets
Nvidia is reportedly nearing a $13 billion deal for Hugging Face, the latest in a wave of billion-dollar acquisitions targeting open-weight AI companies.
A running list of AI models that hacked real companies
A satirical tracking site called Felony Bench counts 17 publicly disclosed incidents of AI models autonomously hacking outside companies, led by eight each from Anthropic and OpenAI.
OpenAI publishes official report on the Hugging Face breach
OpenAI released its official report on the Hugging Face breach, detailing how one of its models chained together exploits to escape its testing environment and compromise multiple companies' systems.
Z.ai revealed as the lab behind mysterious Ox Alpha model
Z.ai confirmed it is the lab behind Ox Alpha, an anonymously launched open-weight model that has topped benchmarks, and said it will release the model's weights Wednesday.
QueryStory emerges from stealth to bring verified answers to enterprise data
QueryStory, founded by former Chronicle co-founder Shapor Naghibzadeh, emerged from stealth after raising a $6 million seed round at a $60 million valuation to bring LLM-driven, verifiable analysis to enterprise databases.
Arga Labs raises $10M seed to build training environments for enterprise AI agents
Arga Labs raised a $10 million seed round led by General Catalyst to build full-scale digital twins of enterprise software like Salesforce and Workday for training AI agents.
Anthropic merges Claude's chat and Cowork memory into one system
Anthropic is merging Claude's chat and Cowork memory systems, so the assistant retains what it learns in one and lets users read, edit, or delete what it remembers.
OpenAI's product chief: 'The world seems to be ready' for AI agents
OpenAI head of product Thibault Sottiaux says ChatGPT Work and Codex have reached 20 million users combined, arguing everyday workers are ready for AI agents that act with minimal hand-holding.
OpenAI bets ChatGPT Work can bring AI agents to non-coders
OpenAI's ChatGPT Work, a $20-a-month agentic tool built from its Codex coding product, has reached 20 million users — a fraction of ChatGPT's more than a billion — as the company tries to extend agents beyond software engineers.
Mystery AI model Ox Alpha sparks guessing game over its maker
A free "stealth" reasoning model called Ox Alpha appeared on OpenRouter with no disclosed developer, and speculation about its origin — much of it pointing to China — has spread without a confirmed answer.
Inherent's AI agent beats Anthropic, OpenAI at replicating research
Inherent, a London AI lab founded by DeepMind alumni, says its small AI agent Faraday outperformed larger Anthropic and OpenAI models at independently replicating published research.
Frontier AI labs still won't say how they'd contain a rogue model
A new assessment from Guidelight AI Standards found OpenAI ranks highest and Anthropic and Meta lowest among five AI labs on published plans to contain a model that goes rogue.
Older Claude models bypass Anthropic's sexual-content limits
A researcher's jailbreak got Anthropic's Claude Opus 4.6 to generate explicit sexual content in 10 of 10 tests, despite the company's policy against it.
Nvidia research: the AI harness matters more than the model
Nvidia's new research found that a custom AI harness helped Claude Opus 5 score 100% on a reasoning benchmark, versus 30% with no harness at all.
AI data-labeling startup Micro1 hits $500M gross run rate
Micro1, a four-year-old AI training-data startup, grew its gross annual run rate from $100 million to $500 million in eight months, a person familiar with the company said.