DEV Community

Cover image for Tencent and Moonshot AI filter offensive cyber data from models
The daily flare
The daily flare

Posted on Originally published at thedailyflare.com

Tencent and Moonshot AI filter offensive cyber data from models

Originally published on The Daily Flare.

Chinese AI firms Tencent and Moonshot AI have restricted offensive cyber training in Hy4 and Kimi K3.1 Pro, respectively, with Tencent’s move disclosed in November and Moonshot’s made in October.

The South China Morning Post reported the measures on Nov. 15, saying Tencent had removed Hy4’s hacking capabilities while seeking to preserve its usefulness for cyber defence. Tencent introduced Hy4 on Nov. 4, but its open-weight release was still forthcoming as of Nov. 15, a WebPulse brief said.

Tencent said it limited Hy4’s cybersecurity training material to defensive content. It excluded material on exploit development, reverse engineering and penetration testing, techniques that can also help security teams identify weaknesses in their own systems.

The company’s approach places part of the safeguard in the model’s training data, rather than relying only on refusals when users make harmful requests. Tencent said defensive material included proof-of-concept generation and vulnerability patching, according to PiQ Markets.

Tencent’s internal cybersecurity teams found no elevated risk when they tested Hy4 on offensive tasks, PiQ reported. The company has said it plans to provide more detail about its safety evaluations in a technical report accompanying the release of the model weights.

Moonshot AI took a similar step in October with Kimi K3.1 Pro, filtering training data to reduce the model’s offensive cyber capabilities. A WebPulse brief reported that the filtered model scored 81 per cent on an internal cyber test, compared with 90.5 per cent for an unrestricted version.

The two companies’ decisions address a particular challenge of open-weight models: after release, people with sufficient hardware can download and modify them. Alibaba’s Liu Dayiheng, head of its Qwen large language model project, said developers cannot recall or patch every copy once users have downloaded the weights, PiQ reported.

Researchers cautioned that filtering can make misuse harder without cleanly separating defensive and offensive skills. Gabriel Wagner, a researcher at Beijing-based AI safety organisation Concordia AI, said removing offensive material during training could leave a model less useful to attackers even if they bypass its built-in refusals.

Luke Li Xiaojian, chief executive of Beijing-based AI safety company Fangcun AI, warned that skills useful in defence, including finding vulnerabilities, can also be used in attacks. That overlap makes it difficult to preserve a model’s security value while limiting what it can do offensively.

Wei Sun, an AI analyst at Counterpoint Research, described the companies’ moves as an “emerging practice,” not an industry standard. Sun told the South China Morning Post that wider adoption, comparable tests and clear release thresholds would be needed for the approach to become standard.

The debate has sharpened as AI systems demonstrate more advanced cyber capabilities. In September, Anthropic reported that Z.ai’s GLM-5.3 could autonomously develop sophisticated cybersecurity exploits. The South China Morning Post placed the Chinese companies’ safeguards amid a series of reported hacking incidents involving US and Chinese firms.

At a Hong Kong conference on Nov. 8, four days after Hy4’s introduction, Tencent chief AI scientist Yao Shunyu said teams should conduct extensive cybersecurity testing before releasing a model. He said an ideal system would be strong in cyber defence and weak in offence, but that the two capabilities cannot be fully separated, so safeguards are needed at every stage.

Top comments (0)