All articles
Midjourney vs. Hollywood: Prove you use AI yourselves
In its copyright dispute with Disney, Universal, and Warner Bros., Midjourney has moved to compel the studios to disclose their internal AI practices. The startup argues that if the studios themselves train models on copyrighted content, their lawsuit is hypocrisy.
Anthropic in talks with Samsung on a custom AI chip
Anthropic is in early talks with Samsung to co-develop a custom AI chip, responding to OpenAI's Jalapeño inference processor announced with Broadcom a week earlier.
Cloudflare will block AI crawlers mixing search and training by default from September 15
Cloudflare announced that from September 15, 2026, mixed-use AI crawlers will be blocked by default on ad-supported sites. The company is also expanding Pay Per Crawl into Pay Per Use: billing AI companies when their content actually generates value, not just when it is fetched.
New attack on AI browsers: telling the LLM 2+2=5 bypasses safety guardrails
Researchers demonstrated that feeding an LLM-based browser a false fact is enough to make it abandon its safety guardrails. The BioShocking attack worked on six AI browsers.
X launches an MCP server — platform data now available to AI agents
X (formerly Twitter) launched a hosted MCP server that lets AI tools like Claude and Cursor access platform data through a user's own account — without developers having to build custom integrations.
Anthropic restores Fable 5 and Mythos 5 globally with a new jailbreak severity framework
Anthropic restored global access to Claude Fable 5 and Mythos 5 on July 1, 2026, after the US lifted export controls in place since June 12. The company also published a proposed industry-wide jailbreak severity framework co-developed with Amazon, Microsoft, and Google.
Meituan open-sources LongCat-2.0: 1.6T MoE on Chinese ASICs beats GPT-5.5
Meituan unveiled LongCat-2.0, a 1.6T-parameter MoE trained without Nvidia GPUs, which spent two months anonymously leading OpenRouter as "Owl Alpha". It scored 59.5 on SWE-bench Pro, above GPT-5.5.
South Korea commits $1T: memory chips, AI data centers, and humanoid robots by 2028
South Korea announced $1T in megaprojects: Samsung and SK Hynix build new DRAM fabs ($585B), Naver and SK Group invest in AI data centers ($357B), and Hyundai commits $5.8B to an Atlas robot factory. The target is 30,000 humanoids per year by 2028.
Claude Science: Anthropic launches an AI research workbench for scientists
Anthropic launched Claude Science, a research workbench for scientists integrating 60+ databases, multi-agent workflows, and citation verification. It is not a new model — the platform runs on existing Claude Opus 4.8. Beta available from June 30, 2026 for subscribers.
Anthropic releases Claude Sonnet 5 — agentic power at Sonnet 4.6 price
Anthropic released Claude Sonnet 5, a new mid-tier model with substantially improved agentic capabilities and performance close to Opus 4.8 at introductory pricing of $2/MTok input.
morph's soft robotic cells: physical AI built into the material
London startup morph unveiled a platform of soft robotic cells with embedded intelligence — the cells change shape and stiffness in real time in response to environmental signals. The company supplies the technology as a component for robot manufacturers.
When should AI speak up? Tsinghua releases two proactive-AI benchmarks
Two teams from Tsinghua University published the EgoIntrospect and IPIBench benchmarks in May 2026, measuring how well multimodal large language models understand users' internal states and act proactively in real time. No tested model passed both evaluations.