Anthropic sees growing potential for language models in cybersecurity. The company cites results from the CyberGym leaderboard: Claude Sonnet 4 uncovers new software vulnerabilities about 2 percent of the time, while Sonnet 4.5 increases that rate to 5 percent. In repeated tests, Sonnet 4.5 finds new vulnerabilities in more than a third of projects.

In a recent DARPA AI Cyber Challenge, Anthropic notes that teams used large language models like Claude "to build 'cyber reasoning systems' that examined millions of lines of code for vulnerabilities to patch." Anthropic calls this a possible "inflection point for AI’s impact on cybersecurity."
Meta's top AI researcher, Yann LeCun, is reportedly at odds with the company over new publication guidelines for its FAIR research division. According to six people familiar with the matter, FAIR projects now need stricter internal review before release - a shift some employees say limits their scientific freedom. LeCun even considered stepping down in September, The Information reports, partly in response to Shengjia Zhao being named chief scientist for Meta's superintelligence labs.
The dispute comes as Meta reshapes its AI organization. LeCun, who has openly rejected the current large language model (LLM) paradigm, is pushing for new directions in AI. He has also positioned himself against Donald Trump, while CEO Mark Zuckerberg has been more willing to align with the Trump administration.
OpenAI's Sora 2 can handle knowledge questions, too. In a test by Epoch AI, Sora got ten random tasks from the GPQA Diamond Multiple Choice Benchmark covering natural sciences. Sora scored 55 percent, while GPT-5 managed 72 percent. To run the test, Epoch AI asked Sora to make a video of a professor holding up the answer letter on a sheet of paper.
Video: via EpochAI
Epoch AI points out that an upstream language model could tweak the prompt before the video is created and slot in the answer along the way. Other systems, like HunyuanVideo, use similar re-prompting tricks, but it's not confirmed whether Sora does the same. Either way, the lines between text and video models are starting to blur.
Meta is set to start mining users' conversations with its chatbot to target ads and content across all Meta platforms, including Facebook and Instagram. Beginning December 16, 2025, anything users say to Meta AI—by text or voice—will feed into the company's ad and content algorithms. If someone discusses hiking with the AI, for instance, they can expect to see more hiking-related ads, posts, and groups. Meta says it will exclude sensitive subjects like religion, health, and political views from this data collection.

Users can try to limit what shows up in their feeds using settings like "Ads Preferences," but these changes only apply if accounts are linked in the Accounts Center. Meta plans to alert affected users ahead of time by notification and email. The policy will take effect in most regions.
OpenAI has acquired the AI finance startup Roi, known for offering investment advice based on individual financial data. The service will shut down on October 15. Roi's CEO and founder Sujith Vishwajith is the only team member joining OpenAI. The deal supports OpenAI's interest in personalized consumer AI. The purchase price was not disclosed.

ChatGPT still dominates the AI market, but Google Gemini is catching up. Over the past year, Similarweb says Gemini's share of generative AI traffic jumped from 6.5 percent to 13.7 percent. ChatGPT leads at 73.8 percent, down from 87.1 percent a year ago.

The rest of the market is much smaller. DeepSeek holds 3.9 percent, while Perplexity and Grok each have 2.0 percent. Claude sits at 1.8 percent, and Microsoft's Copilot brings up the rear with 1.2 percent. These numbers haven't shifted much for the smaller players. Overall traffic to AI services continues to grow.
Alibaba's Qwen group has released two new small-scale multimodal models, Qwen3-VL-30B-A3B-Instruct and Qwen3-VL-30B-A3B-Thinking, each with 3 billion active parameters. According to Qwen, both versions are competitive with GPT-5-Mini and Claude 4 Sonnet, and in some benchmarks show stronger performance in math, image recognition, text recognition, video processing, and agent control.
The lineup includes an FP8 version for faster inference, and an FP8 variant of the Qwen3-VL-235B-A22B model. The models are available on HuggingFace, ModelScope, and GitHub, or via an Alibaba Cloud API. There is also a web chat interface for direct use.