BEIJING — Tencent on Wednesday, November 4, 2026, released Hy4, its new open-source artificial intelligence model, trained with deliberately limited cybersecurity data so that it leans toward "defensive" uses rather than helping cyberattacks, Reuters reported — the latest sign that AI developers and governments are moving, in the same week, to contain a wave of AI-driven hacking.
Illustration: A digital shield guards computer servers as Tencent says its new Hy4 model was trained to limit its ability to aid cyberattacks.
What Happened
Tencent said it filtered out content during training such as exploit development, reverse engineering and penetration testing to limit Hy4's potential "offensive" capabilities, according to Reuters. The company said the model was trained with limited cybersecurity data, restricted to "defensive" content, in an effort to reduce the risk that it could be used to aid attacks.
"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," Tencent said in a statement reported by Reuters.
The rollout comes after a slew of AI-driven cyberattacks around the world raised concerns about the capabilities of advanced models and the threat they pose to global internet infrastructure, Reuters reported. Tencent is not the first Chinese developer to take this step: Moonshot AI last month limited its newest model's cyber-offensive capabilities, as Chinese AI firms join US companies in trying to prevent loss-of-control risks, according to Reuters.
The same day brought warnings and rule-making on three continents. India's cybercrime agency issued an advisory to financial firms, including banks and non-bank lenders, about cyberattacks carried out with AI agents, days after the country's central bank issued a similar warning. In the United States, the Office of the Comptroller of the Currency issued what it said was its first-ever guidance for how bank examiners should probe lenders' AI use. And Japan plans to seek earlier access to information on undisclosed software vulnerabilities from US artificial intelligence developers and other overseas technology firms, the Nikkei newspaper reported, as it strengthens defences against a rising tide of cyberattacks.
Background
The debate over AI and hacking sharpened in October. Moonshot AI announced on October 24 that it had reduced the cyber-offensive capabilities of its newest model, Kimi K3.1 Pro, which it describes as its most capable 3-trillion-parameter-class open-weight model, Reuters reported. The Beijing-based startup said it removed cyber-related data from the training corpus and from reinforcement-learning environments used to develop the model, while retaining general coding skills.
"We aim not to release open weights that would extend the absolute frontier of cyber capability," Moonshot said, according to Reuters. The company said Kimi K3.1 Pro scored 81.0% on its internal cyber benchmark, compared with 90.5% for a similar version without the risk-reduction measures, although those results have not been independently verified. It planned to publish the model's weights — the files that allow outside developers to run and modify it on their own systems — by November 4.
The timing matters because open-weight releases are difficult to reverse. Once weights are public, any developer can download, adapt and run the model without the original developer's controls. That is why both Moonshot's and Tencent's decisions were made before release, at the training stage, rather than as filters bolted on afterwards.
October also exposed how fragile widely used software can be. Government and corporate websites in several countries were disrupted in October after the disclosure of critical vulnerabilities in the PHP programming language that underpins much of the internet, prompting millions of hacking attempts and forcing website operators to protect or temporarily shut down their systems, Reuters reported, citing Nikkei. As organisations scrambled to mitigate the PHP risks, Japan said it would set up a team to simulate cyberattacks using AI to identify weaknesses in technology systems.
Illustration: Analysts monitor global cyber threats as AI agents are increasingly used to carry out attacks at machine speed.
Why It Matters
The Tencent release matters because it tests whether the AI industry can police a dual-use problem on its own. The same skills that make a model useful for defending software — finding flaws, understanding exploits, writing and reviewing code — can also be turned to attacking it. By removing exploit development, reverse engineering and penetration-testing material from training, Tencent is accepting a trade-off: a model that may be less capable at some security tasks in exchange for one that is harder to weaponise.
For the financial sector, the stakes are immediate. India's cybercrime agency, the Indian Cybercrime Coordination Centre, said in an advisory made public on Wednesday that attacks on financial firms were financially motivated and carried out with the help of AI agents. "These attacks are financially motivated and are being carried out with the help of Artificial Intelligence (AI) agents leading to machine speed compromise," the centre said, according to Reuters. When attacks move at machine speed, human defenders have less time to react — which is why regulators are now focusing on how much autonomy AI systems are given inside banks.
The US bank regulator's new guidance reflects that shift. The Office of the Comptroller of the Currency, which oversees large national banks, said examiners should monitor how lenders use AI but should not treat it as distinct from other risks, according to a copy of the document seen by Reuters. Banks are not expected to establish separate risk parameters for AI use, but examiners should consider how a lender uses AI and the size and complexity of the lender. While the document is not legally binding like a rule, banks typically comply closely with such guidance, Reuters reported.
Crucially, the guidance tells examiners to focus on the consequences of potential AI-driven mistakes rather than the sophistication of particular models. A simple AI agent authorised to initiate transactions could be considered riskier than a sophisticated model deployed for internal purposes, the document said. Higher-risk uses of AI may require greater restrictions on autonomy or more direct human oversight.
For governments, the week also shows a move from asking AI companies questions to demanding earlier information. Japan had recently sought expanded access to advanced AI models from major US developers, including Anthropic and OpenAI, in response to a wave of cyberattacks believed to have been aided by AI, Nikkei reported. Tokyo will now explore ways to obtain information about serious software vulnerabilities before they are publicly disclosed, so government agencies and operators of critical infrastructure can respond more quickly.
What Happens Next
Japan's government is expected to present an emergency cybersecurity package at a ministerial meeting that could be convened as early as Friday, Nikkei reported. In late October, Prime Minister Sanae Takaichi instructed officials to devise steps to bolster cyber defences following a recent increase in cyberattacks, according to Nikkei. The package will be watched as a test of how far a government can go in asking foreign AI firms for pre-disclosure vulnerability data — information companies have traditionally guarded closely.
In the United States, bank examiners will now apply the OCC framework in their supervision of large lenders, and banking regulators worldwide have been querying lenders about their AI use, worried the technology poses major cybersecurity and potentially financial-stability risks, Reuters reported. In the US, regulators have been asking banks about their use of advanced AI models, what limits banks place on those systems and how they monitor them.
For AI developers, the question is whether Tencent's approach becomes a norm for open-weight releases. Moonshot's Kimi K3.1 Pro weights were due to be published by November 4, and Tencent's Hy4 is now out in the open. Researchers and rival labs will scrutinise both models to judge whether the safety trade-offs actually hold in practice — and whether models trained with less offensive cyber data can still defend systems effectively.
If the curbs work, they could offer a template for future releases: limit the most dangerous training data, keep general coding ability, and publish the reasoning behind the choice. If they do not, regulators in India, Japan, the United States and elsewhere have already shown they are prepared to step in with advisories, guidance and demands for earlier disclosure — and the next round of rules may be less voluntary than this one.
Illustration: A bank building under a digital AI brain, as US regulators issue their first guidance on examining banks' use of artificial intelligence.
Sponsored: explore today's featured sponsored offers
FAQ
What is Tencent's Hy4 model?
Hy4 is Tencent's new open-source AI model, released on November 4, 2026. Tencent said it was trained with limited cybersecurity data, restricted to "defensive" content, to reduce the risk that it could aid cyberattacks, Reuters reported.
What training data did Tencent filter out?
Tencent said it filtered out content such as exploit development, reverse engineering and penetration testing to limit Hy4's potential "offensive" capabilities, according to Reuters.
How does this compare with Moonshot's Kimi K3.1 Pro?
Moonshot AI announced on October 24 that it had reduced the cyber-offensive capabilities of Kimi K3.1 Pro by removing cyber-related data from training, with its internal benchmark score falling to 81.0% from 90.5% for an unfiltered version. Those results have not been independently verified, Reuters reported.
What did India warn financial firms about?
India's cybercrime agency warned banks and non-bank lenders about financially motivated cyberattacks carried out with AI agents, which it said lead to "machine speed compromise". The advisory was made public on November 4, days after the central bank issued a similar warning, Reuters reported.
What is the US bank regulator's new AI guidance?
The Office of the Comptroller of the Currency issued its first-ever guidance for examiners on banks' AI use, saying examiners should focus on the consequences of AI mistakes rather than model sophistication, and that higher-risk uses may need greater restrictions on autonomy or more human oversight. The guidance is not legally binding, Reuters reported.
Sources
Reuters, "Tencent rolls out AI model trained with limited cybersecurity data amid hacking fears", November 4, 2026; Reuters, "India cybercrime agency warns financial firms on AI attacks", November 4, 2026; Reuters, "Japan to seek software vulnerability data from US AI firms amid cyberattack surge", Nikkei report, November 4, 2026; Reuters, "US bank regulator issues first-ever AI guidance for examiners", November 4, 2026; Reuters, "Moonshot curbs hacking ability in new Kimi model as fears grow over AI-powered cyberattacks", October 24, 2026.
Comments
Post a Comment