China's Tencent says it removed hacking lessons from its new Hy4 AI

Exploit writing, reverse engineering, penetration testing: these three attacker skills were removed from the training data of Hy4, the new AI model from Chinese tech giant Tencent, released on Wednesday, November 4. The company says it limited its cybersecurity data to "defensive" content to reduce the risk that its AI is used to carry out cyberattacks, Reuters reports.

"We recognize the potential risks posed by advanced capabilities in cybersecurity and biology," Tencent writes, promising a "more responsible" approach to training and releasing its models. Reuters places the decision in the wake of a series of AI-assisted cyberattacks around the world. Last month, another Chinese company, Moonshot AI, had already curbed the offensive capabilities of its latest model, Kimi K3.1 Pro.

How good is an AI that mostly learned to defend?

According to the European site Trending Topics, Tencent reserves its model's cybersecurity skills for defensive uses, such as spotting and fixing flaws in software. Tencent says experts from its Xuanwu and Keen labs found no notable concerns about the model's offensive capabilities in the areas evaluated, but the company publishes no figures to back this up. For now, it is impossible to verify from the outside what the filtering actually removed.

The line between attack and defense is thin, in any case. Security teams, for example, disassemble malware to protect against it. Tencent says it prioritized writing proofs of concept and fixing flaws, but does not say whether removing reverse engineering hampers these legitimate uses.

An "open" model whose weights are slow to arrive

Reuters describes Hy4 as an open-source model. In practice, the final version is currently available only online: through an API, on Tencent's cloud platform and via the OpenRouter service, as well as in Tencent apps such as its Yuanbao assistant. Its weights, the parameters that let you run the model on your own servers, will be released along with a report on safety testing "in the coming days," according to Tencent's statement, which gives no firm date.

Only a preview version has been downloadable since late August, on Hugging Face, under the Apache 2.0 license, which allows commercial use. It is a so-called "mixture-of-experts" model with 770 billion parameters, 49 billion of which are activated at each step. It can process about one million tokens (word fragments) at once. On pricing, Trending Topics reports 6 yuan (about €0.76) per million input tokens and 18 yuan (about €2.25) for output, at half price until the end of the year.

Performance still to be verified

Tencent reports that its rate of "hallucinations," meaning invented answers, fell from 15.76% to 5.45% compared with the preview version. It adds that Hy4 can analyze images and operate a computer, both from the command line and in on-screen applications. All of these figures come from the developer. On Terminal Bench 4.0, a test of tasks to complete in a terminal, it reports 44.4 points itself, versus 66.4 for Anthropic's Claude Opus 5.5 and 55.5 for OpenAI's GPT-6.1 Sol. Tencent also acknowledges a "notable gap" with the best proprietary models in hard mathematics and long software projects. At launch, no independent evaluation was available: the benchmarking site Artificial Analysis had not yet ranked it.

Before you try it

• Through the API, your requests go through Tencent's infrastructure. For sensitive data, it is better to wait for the final version's weights and host it yourself.
• Check the final version's license when it is published: for now, only the preview's license is known.
• Don't rely on the scores published by the developer: test the model on your own tasks.

Publishing the weights and the safety report will finally let independent researchers measure what Tencent's AI can still do when it comes to hacking.