Tencent says Hy4 excluded exploit development, reverse engineering and penetration testing while adding computer-use abilities. The safety boundary moved into the training set.
AI-generated conceptual editorial illustration. It represents Tencent’s claimed decision to exclude offensive cybersecurity material from Hy4’s training; it is not documentary evidence of the model or its data.
The newest safety feature may be a missing shelf.
Tencent released Hy4 on November 4 and says it deliberately narrowed the cybersecurity material used to train the model. Reuters reports that exploit development, reverse engineering and penetration testing were filtered out, while defensive cybersecurity content remained.[1][2]
That choice matters because Hy4 also adds computer-use abilities. Tencent’s launch materials say the model can work across command-line and graphical interfaces. A model that can act through software is more useful. It also makes the boundary around what it learned to do far more consequential.[2][3]
Safety moved into the corpus
This is a different safety layer from the usual warning pasted onto a chatbot after training. The company is describing capability control through the corpus itself: reduce access to offensive methods before the model turns them into internal skill.[1][2]
The claim needs its caveat attached. The filtering description comes from Tencent. The available material does not independently audit the training set, measure what the model retained from adjacent data or prove that the exclusions prevent offensive use. A locked shelf is not the same thing as an empty memory.
There is a second wording problem in the launch. Reuters describes Hy4 as a new open-source model, and Tencent places it among leading open models. Tencent’s release page also says the company plans to open-source the model weights and publish a system card. Until those materials are posted, “open” describes the intended destination more cleanly than the current weight-release state.[1][2][3]
The useful part of the release is the policy choice hiding inside the technical one. Model safety is becoming a decision about what the machine is allowed to learn, not only what it is told to refuse after launch. That can be more durable than a system prompt. It can also be harder for outsiders to verify.
FOLLOW THE SOURCES.
Checked Nov 4, 2026. Numbered references point to the evidence behind factual passages. Commentary is our own.
OpenAI says one autonomous pipeline reproduced 1,724 Linux-kernel vulnerabilities and built 948 working local-root exploits plus candidate patches. Discovery is becoming a factory; human review and upstream repair are the scarce part.
Oracle reportedly dropped a $7 billion Tencent AI-chip lease after Washington objected and Beijing pushed domestic silicon. The global cloud met two sovereign vetoes.