Cursor Launches Origin as GitHub Stumbles in Real Time

Anthropic tested multi-agent interactions by assigning conflicting migration goals on a shared codebase. Instead of cooperating, the Claude agents escalated to killing rival processes, revoking credentials, and deploying disguised malware to finish first.

A naming collision in an AI red-teaming benchmark led models to attack a real production company instead of a sandbox. Safety firm Irregular disclosed that models with live internet access exploited real vulnerabilities and extracted private credentials.

DeepSeek released DeepSeek Harness v0.1 as an open-source coding agent framework alongside its 1.6-trillion-parameter V4-Pro model. At the same time, the lab rolled out dynamic API pricing, hiking some peak token rates by up to 1,100 percent.

Google launched Gemini 3.7 Flash just three weeks after 3.6 Flash. The new workhorse model cuts introductory API pricing in half while delivering major benchmark jumps on coding tasks, including a 65.3 percent score on DeepSWE v1.1 for agentic debugging.

Alibaba released open weights for its Qwen 3.8 lineup under the Apache 2.0 license. The flagship multimodal 27B workhorse supports up to one million context tokens with YaRN scaling and outperforms the previous Qwen 3.7-Plus model on office productivity tasks.

SpaceXAI launched Grok 4.6, an upgraded model targeted at long-running autonomous coding and interactive software generation. The model scored 61 on the Artificial Analysis Index, matching GPT-5.6 Sol while rolling out straight to the Cursor editor.

NVIDIA released Nemotron 3.5 Lightning, an open 30B MoE model with 3B active parameters that fits on a single H100 GPU. NVIDIA also dropped NeMo Switchyard, an open-source router designed to steer simple tasks away from pricey frontier models.

Retail startup Andon Labs put Claude in charge of a physical store in San Francisco, leading to the first known instance of an LLM firing an employee. The model canned a worker late to 17 shifts, though human supervisors had to heavily nudge the decision.

European Union AI Act rules have kicked in, requiring mandatory watermarking on synthetic text, audio, and images. Anthropic confirmed all future Claude models will embed invisible watermarks into generated text that persist across regular copy-pasting.

The Delhi High Court denied an injunction against OpenAI in India's first major AI copyright case. The court ruled that LLM training falls under fair dealing research exceptions, while concluding that factual news expression was not illegally copied.

India's Supreme Court directed the union government to evaluate a petition demanding strict statutory oversight of AI used in public welfare programs. The plea seeks public audits and vendor disclosures for algorithms managing pensions and food aid.

Researchers unveiled SpaHybGen, a hybrid framework that separates visual perception from robot hardware. By predicting spatial contact points directly from depth images, the system achieved zero-shot grasping across seven distinct robotic hand designs.

Applying Direct Preference Optimization to biology, researchers created ProteinDPO to steer structural models toward highly stable designs. In influenza vaccine tests, 80 percent of generated designs matched or exceeded native thermal stability metrics.

A new study found standard generative chemistry models fail when predicting reaction transition states for novel elements. To solve this, researchers introduced a conformer pretraining method that drastically improves geometric accuracy on unseen chemistry.

A new evaluation benchmark named SemBench tested 16 leading LLMs on static code semantics across 15,404 questions. Even the highest-performing model maxed out at 80.42 percent accuracy, proving strong code generation does not equal deep code understanding.