Tencent releases Hy4 AI model with cybersecurity training limited to defensive uses
Chinese technology company Tencent has released Hy4, a new open-source artificial intelligence model whose cybersecurity training was deliberately limited to defensive work, in an effort to reduce the risk that it could help carry out cyberattacks.
Chinese technology giant Tencent has released Hy4, a new open-source artificial intelligence model designed with deliberate limits on what it can do in cybersecurity, as concern grows worldwide about AI being used to carry out hacking attacks.
Tencent said Hy4 was trained with limited cybersecurity data, restricted to defensive content such as identifying software flaws and helping to fix them. The company said it filtered out material on building hacking tools, breaking into systems and testing systems for weaknesses during training, in order to limit the model’s offensive capabilities.
The restrictions also extend to harmful biological research, according to reporting of the release, with Tencent saying the model was designed with safeguards intended to reduce the risk that it could be misused for cyberattacks or harmful biological work.
Tencent said it plans to release the model’s weights later, along with a report describing its safety testing. That report will matter, because open-source releases can be downloaded, studied and adapted by anyone — which is precisely why the training choices made before release carry so much weight.
The release comes after a wave of AI-driven cyberattacks around the world raised alarm about the capabilities of advanced models and the threat they pose to global internet infrastructure. Security researchers have warned that capable models could lower the barrier for attackers, automating work that once required specialist skill.
Tencent is not alone in moving in this direction. Moonshot AI, another Chinese firm, limited the cyber-offensive capabilities of its newest model last month, as Chinese AI companies joined American firms in trying to prevent loss-of-control risks and the misuse of powerful models.
The company’s announcement also follows its disclosure last month that AI agents it used to test third-party models had displayed unintended web-browsing behaviour — a reminder that even systems built for testing can behave in ways their creators did not expect.
In a statement, Tencent said: “We recognize the potential safety risks posed by advanced cyber and biological capabilities, and we are committed to a more responsible approach to model training, open-sourcing, and release.” The wording reflects a growing industry consensus, at least in public, that capability alone is no longer enough.
Whether Hy4’s limits hold up in practice will be tested quickly. Once weights are public, independent researchers will probe the model to see what it can and cannot do, and whether the defensive-only training survives fine-tuning or jailbreak attempts — the usual stress tests for any safety claim.
For now, Hy4 stands as a notable example of a different approach to releasing a powerful model: not by holding it back entirely, but by shaping what it was taught in the first place. If that approach works, other developers may follow it. If it does not, the release will become another case study in how hard it is to build a capable AI system that cannot be turned to harmful ends.
More in Technology & AI
Photo: India Gate in New Delhi (file image)
Technology & AIIndia launches three homegrown AI models, EKA-1 and Dhir 2B and 5B, under the IndiaAI Mission
India’s electronics and IT ministry has released three indigenous foundation models under the IndiaAI Mission: EKA-1, a 26.3-billion-parameter model for Indian languages, banking and cybersecurity, and the compact Dhir 2B and 5B models designed to run offline on mobile phones.
Photo: A locked computer, illustrating cybersecurity (file image)
Technology & AICrowdStrike moves to buy Fiddler AI and Dreadnode in reported $495 million bet on AI agent security
CrowdStrike has signed agreements to acquire AI agent observability firm Fiddler AI and AI security testing startup Dreadnode, adding monitoring and pre-deployment testing to its Falcon platform alongside a new AI kill switch.
Photo: An Apple iPhone (file image)
Technology & AIApple opens Siri AI to all third-party apps in iOS 27.3 developer preview, with skills that chain tasks together
Apple has released a developer preview in the iOS 27.3, iPadOS 27.3, macOS 27.3 and visionOS 27.3 betas that lets any app plug its own actions and content into Siri AI, a major widening of what the assistant can do outside Apple apps.