Recently, the "Godfather of AI" and Nobel laureate in Physics Geoffrey Hinton has once again issued a warning. In his view, there are a series of security risks in AI development. For instance, merely launching large-scale cyberattacks and creating cybersecurity viruses could be enough to cause severe chaos.
This is not science fiction, but a serious proposition facing the AI community. When agents take over business processes and attackers pick up AI tools, cybersecurity is no longer just a "cost item" for enterprise IT, but has escalated to industrial security, social security, and even the common security of humanity. Recently, some overseas AI enterprise executives have also called for slowing down the pace of AI R&D, with this very issue being the core consideration.
During Huawei Connect 2026, Wu Jiaxing, President of the Security Product Domain of Huawei's Data Communication Product Line, accepted a media interview and also focused on this topic. Facing the unprecedented imbalance between attack and defense in the AI era, Huawei's answer is not to slow down AI, but to "use AI to protect AI": on the one hand, defining and promoting the AI Firewall (AIFW) standard to lay out the next-generation security architecture; on the other hand, deploying the Xinghe AI three-tier security fence deep into computing centers to bring security capabilities to the front line of production.
Wu Jiaxing, President of the Security Product Domain of Huawei's Data Communication Product Line
Both Defense and Attack Logics Have Changed; Security Must Adopt a New Approach
"The main changes in security in the AI era come from two aspects: one is that the enterprise's defense business logic has changed, and the other is that the hackers' attack logic has changed," Wu Jiaxing pointed out straightforwardly.
In the past, network entities were relatively clear—humans accessed applications, and firewalls identified "humans and applications." Today, with the explosion of AI Agents, non-human identities have become important initiators of traffic—model invocations, tool executions, and automated workflows shuttle in and out of enterprises, extending the protection targets to large model data and automated processes. This means that the security boundary is no longer just a "wall" at the gate of the campus, but a full-link governance running through users, applications, models, agents, and infrastructure.
The changes on the attack side are even more direct. AI significantly lowers the threshold for attacks and also drives viruses to iterate and mutate at a speed far beyond manual analysis. Wu Jiaxing summarized the dilemma of traditional security devices in one sentence: "Traditional rule-based devices cannot read the attack intentions behind the traffic, and it is also difficult to defend against such new types of attacks."
Therefore, the old system relying solely on "rule base + cloud detection" is simultaneously facing a triple failure: rule updates cannot keep up with attack mutations, cloud round-trip latency cannot stop real-time intrusions, and single-point devices cannot clearly see the semantic behavior of models and agents. Cybersecurity defense must adopt a new approach.
Huawei's approach and practice is to completely change the attack and defense relationship of "humans managing rules and rules chasing viruses," and use "intelligence to counter intelligence." Wu Jiaxing stated, "Our consistent approach is still to use AI to counter AI, driving security from rule-driven to AI-driven, from perception to intention, and from manual operation to autonomous defense."
AIFW Standards and Specifications: Turning "Everyone Speaking Their Own Language" into Industry Consensus
Previously, some vendors put the concept first, calling traditional firewalls with AI detection capabilities AIFWs. But what the industry needs is a new generation, not just stacking a few AI functions on old products. Without a "common language" and a unified selection scale for customers, it is difficult for the industry to scale. To this end, Wu Jiaxing shared that Huawei has mainly done work on two levels.
At the national standard research level, under the guidance of the National Information Security Standardization Technical Committee (TC260), Huawei took the lead in carrying out AI cybersecurity technology standardization research, sorting out the changes in attack patterns brought by AI and the capability gaps of traditional security products, covering product systems such as AIFW, AIDR, and AISOC.
At the industry specification level, Huawei joined hands with the China Cybersecurity Industry Alliance (CCIA) and more than 20 mainstream vendors and research institutions to jointly compile the country's first AIFW technical specification, "Security Technical Requirements for Artificial Intelligence Firewalls." Building on GB/T 20281, this specification adds exclusive security indicators for AI models, agents, and tool invocations, providing a unified scale for industry product R&D, testing, and implementation, and ending the conceptual melee.
Where exactly is the "newness" of AIFW? Wu Jiaxing pointed out three key differences. First, the protection targets are extended from "humans + applications" to full-dimensional governance of users, applications, models, agents, and automated workflows. Second, hardware-level AI Core local inference replaces cloud detection, supporting online inference locally to efficiently block unknown threats. Finally, intention-driven policy autonomy replaces manual rule operation, achieving automated handling of 99% of security incidents.
Regarding taking the lead in unifying AIFW standards and specifications, does it mean Huawei wants to make its own route the only answer in the industry? Wu Jiaxing specially emphasized, "The purpose of setting standards is definitely not to be exclusive. Security is an ecological industry, and no single vendor can do it all independently. Therefore, we need to build a public stage to avoid reinventing the wheel and enable more partners to jointly grow the market."
Landing the Three-Tier Security Fence: Embedding Security into the AI Computing Power Production Chain
If standards and specifications solve industry consensus, then customers today also need products that can be directly deployed to cope with the "future is here" AI attack and defense. In the dialogue, Wu Jiaxing told GeekNet that, based on deep insights and its own accumulation, Huawei has identified three major risk challenges in cybersecurity in the AI era, and created the Xinghe AI three-tier security fence solution centered on AIFW around the AI computing center scenario.
In terms of risks, the first type is AI computing power hijacking, where attackers steal enterprise computing power for "mining," turning expensive Token production into a source of profit for others, which has become the top threat in Token operations. The second type is prompt injection, where hackers bypass traditional detection through carefully constructed natural language to induce large models to leak sensitive data. The third type is model poisoning and data theft; once LLM servers or open-source frameworks are invaded, core models, enterprise data, and even automated workflows may fall.
Once the above risks occur, the price is extremely heavy. Public reports show that a Mexican AI startup team suffered computing power hijacking; the theft of the Gemini API key generated an abnormal computing power bill of about USD 82,000 in 48 hours, pushing the team to the brink of bankruptcy. DeFi enterprise Resolv Labs suffered a prompt injection attack resulting in credential leakage, with direct losses of about USD 23 million. After the LiteLLM open-source framework was poisoned, data from about 500,000 devices of thousands of enterprises were stolen, affecting over 2,000 enterprises including NVIDIA and Samsung.
In terms of solutions, Huawei's Xinghe AI three-tier security fence solution starts from the three-tier protection of network, agent, and host, giving full play to Huawei's AI-native active defense, attack intention perception, and policy autonomous operation and maintenance capabilities accumulated from long-term cultivation in the fields of security and AI, realizing the use of AI to guard AI and ensuring the safe production of Tokens.
The first tier is the network fence, which has a built-in mining detection model with a detection rate of up to 95%. Relying on the hardware-level AI Core to complete online inference locally, it can implement millisecond-level blocking of malicious traffic. Compared with the latency of several minutes for cloud detection, this determines whether many attacks can be blocked at the door at the first time.
The second tier is the agent fence, which has a built-in large model for prompt injection detection with a detection accuracy of up to 95%. Based on Huawei's exclusive Ascend operator optimization capability, it can reduce the detection latency to below 150 milliseconds. The industry average is 500 milliseconds, and when coupled with the 10 to 100 milliseconds inference latency of the large model itself, it will cause obvious user lag.
The third tier is the host fence, which is deployed on LLM hosts to monitor malicious processes in real time, preventing data theft and model file tampering. It must be known that computing power in data centers is extremely precious; if security consumption takes up a little more, the available computing power for business will be less. The CPU occupancy rate of this solution is less than 1%, far below the industry average of 5%, which is equivalent to returning more computing power to Token production.
These three tiers of fences are not a simple stacking of three products, but form a complete defense architecture while performing their respective duties. Through unified semantic linkage handling, they can achieve a defense effect of 1+1+1 > 3.
Conclusion:
In response to the voices of some overseas vendor executives calling to "slow down or even suspend AI development," Wu Jiaxing does not agree with opposing security and development. In his view, the feasible path is for responsible vendors to embed security capabilities into the technology foundation, using AI to counter AI and using AI to protect AI, rather than "patching a security patch" after innovation.
"Every industrial technology leap will bring new security issues. But looking back at history, the challenges brought by security have not destroyed the industry, but have effectively promoted and elevated the development of the entire industry," Wu Jiaxing emphasized. "Risks will not disappear because we slow down our pace; instead, we must actively respond to them. When the material of the spear changes, the material of the shield must also be upgraded."