AI

Hy4 Sees the Screen, CrowdStrike Watches the Agents, and Europe Draws a Copyright Line

By Crypt0 · Nov 4, 2026

Control became the product on Wednesday, and four separate stories asked the same question from four directions. Tencent shipped a model that can see a screen and operate a computer. CrowdStrike moved to buy the tools that watch agents and test them before deployment. A European court adviser drew a line around news retrieval in chatbot answers. Anthropic saw a senior product leader for Claude Code and Cowork announce her departure. Capability keeps expanding, and measurement, governance, ownership, and continuity expand with it.

Tencent launched the full version of Hy4 on November 4, adding vision and computer use to a model family previewed in August. Hy4 can read fine grained image detail and complex charts, reason across multiple images, and operate through command line tools and graphical interfaces. Tencent trained it on professional software such as Blender and CAD tools and on long running tasks spanning several applications, with gains over the preview in instruction following, long document understanding, and deep research. Hy4 is available through Tencent Cloud and OpenRouter and inside CodeBuddy, WorkBuddy, Yuanbao, and ima, with listed API pricing unchanged from the preview at 0.834 dollars per million input tokens, 2.501 dollars per million output tokens, and 0.042 dollars per million cached input tokens.

Benchmarks from Tencent place Hy4 close to leading closed models on computer use while leaving a clear gap on harder work. On OSWorld 2.1, a test of operating a computer, Hy4 scored 78.9, compared with 81.8 for Claude Opus 5.5 in Tencent table. On Terminal Bench 4.0, Hy4 scored 44.4, compared with 66.4 for Claude Opus 5.5. In a company run finance evaluation with 213 finance professionals across 579 real world tasks, Hy4 scored 3.04, ahead of DeepSeek V4.1 Flash at 2.97 and Kimi K3.1 Pro and GLM 5.3 at 2.68 each. Tencent also reports an internal hallucination rate falling from 15.76 percent for the preview to 5.45 percent for the full release. These figures are vendor reported, awaiting independent reproduction. Model weights and a comprehensive system card are planned for a later release, so hosted access is the product today.

CrowdStrike announced on November 3 at Fal.Con Europe that it agreed to acquire Fiddler AI and Dreadnode, two startups covering opposite ends of the agent lifecycle. Fiddler AI covers observability, evaluation, tracing, and behavioral monitoring for agents in production, and CrowdStrike plans to sell it as Falcon Observe. Dreadnode covers automated security testing for models, applications, agents, and Model Context Protocol servers before deployment, and CrowdStrike plans to sell it as Automated AI Security Testing, with the Dreadnode team joining the Cyber Superintelligence Lab. CrowdStrike kept prices private in its announcements, while press reporting put the combined figure near 495 million dollars, about 400 million dollars for Fiddler AI and about 95 million dollars for Dreadnode. Both deals await closing in the fiscal fourth quarter of 2027, subject to customary conditions, with general availability targeted for the same quarter. Alongside the deals, CrowdStrike added an AI Kill Switch to Falcon Guardian, available now, letting an authorized administrator stop an agent, prevent it from relaunching, and cut its network access.

Europe added the legal frame for the same governance question. Advocate General Maciej Szpunar delivered an opinion on November 3 in Case C 250/25, Like Company v Google Ireland, a case the Court of Justice calls the first generative AI case to reach the Court. The opinion is advisory, judges are now deliberating, and judgment arrives at a later date. Szpunar proposed that retrieval and reuse in real time of online press publications by AI chatbots to generate answers may, in certain circumstances, infringe press publisher rights. The proposed line turns on access and substitution, where an answer infringes when the chatbot reached the article unlawfully, for example by bypassing a paywall or anti scraping control, or when recognizable expression in the answer substitutes entirely for the original publication so a reader can dispense with opening the original. Facts alone stay distinct from protected expression, and the adviser recommended treating the training questions as inadmissible in this case. Builders of news search and research assistants now have a clear audit pair, lawful fetch plus downstream reuse, with attribution standing apart from permission.

Leadership continuity entered the same conversation at Anthropic. Claude Code and Cowork product leader Cat Wu announced her exit on November 3, saying her last day was the preceding Friday, according to India Today reporting summarized in the daily AI brief. Her announcement kept reason and next destination private. The departure follows earlier exits by Claude Code contributor Thariq Shihipar and desktop developer Anthony Morris, giving enterprise customers a practical diligence item, clear escalation contacts and roadmap commitments as responsibility changes hands. Portable evaluations and development workflows gain value in exactly this kind of transition, because they let a team measure a tool on its own work regardless of who owns the roadmap.

For readers buying, building, or governing AI this week, the practical read runs through all four stories at once. Test Hy4 on representative workflows before granting write access, demand integration milestones and exportable traces from security vendors, audit retrieval permissions and output substitution together, and keep evaluations portable as product teams change. Capability announcements arrive daily, and teams that measure behavior, document ownership, and preserve choice across vendors will get the most from each new model.

Sources

• Artificial Brief daily AI news
• Unite AI Hy4 report
• Futurum CrowdStrike report
• IPWatchdog copyright report

New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.