SHEYU.AI舍予基业
AI Model Watch

AI is starting to find flaws on its own: who will your business be visible to?

OpenAI's GPT-6 Cyber reproduced 1,724 Linux kernel vulnerabilities, a reminder that AI now finishes whole tasks — and that brands must be visible in its answers.

·3 min read
AI is starting to find flaws on its own: who will your business be visible to?
AI-generated illustration, not a news photograph

OpenAI has published an official introduction to GPT‑6 Cyber, titled "GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents." One set of figures in that material deserves a pause from anyone running a business: under the same research framework, GPT‑5.6 Sol reproduced and exploited 166 Linux kernel vulnerabilities; GPT‑6 Cyber reproduced 1,724 and wrote working exploit code for 948 of them. OpenAI says the findings were disclosed to maintainers in the manner of the Linux community, with candidate patches attached.

The significance for readers lies outside the security field. It shows that AI is no longer only the thing that writes your copy or makes your posters. It has begun to act on its own, carrying a task from start to finish — finding the problem, verifying it, and delivering a solution.

When AI moves from giving advice to handing over results, the gap between companies stops being who knows more and becomes who can get AI to finish the work for them.

Why this shift matters to owners

For the past two years, most people have used AI one question at a time: draft a social post, generate an image, tidy a paragraph. That use is real, but it has a ceiling — every step needs a human to push it.

GPT‑6 Cyber shows another form. OpenAI says its own teams used the model to harden the infrastructure on which frontier models are trained. In controlled red-team testing, an agent using GPT‑6 Cyber chained several vulnerabilities together, escaped a microVM and executed code on the host machine. The team then patched the software, tightened network restrictions and reported the issue to maintainers and vendors.

The point is not that AI can attack. It is that AI can complete a chain independently. Finding, reproducing, generating a fix and delivering the result once took a person or a team weeks. That chain has been compressed.

OpenAI also sets out the boundaries: API customers using zero data retention generally need Private Safety Processing, while standard retention remains the default. Capability is being opened up, but how data is handled comes with conditions rather than unlimited use.

Your customers are already using AI to decide. Are you in the answer?

Bring the view back to your own business. Your customers may not care about Linux kernel vulnerabilities, but they are already asking tools such as Doubao, DeepSeek and Kimi which option is worth trying and which company does something well. The SHEYU AGENT website states it plainly: more and more people no longer scroll through search results but ask AI directly, and if you are not in AI's answer, customers cannot see you.

This is not alarmism; it is a migration of habit already under way. When AI can finish a whole task for someone, people lean more on the answer it gives. Whether your brand is inside that answer has become a new operating question.

What SHEYU AGENT can and cannot do here

If you want to take being seen by AI seriously, two capabilities in SHEYU AGENT bear directly on this.

The first is GEO monitoring. It watches six AI engines — Doubao, Wenxin, Kimi, Tongyi, DeepSeek and Yuanbao — to see whether you are mentioned or recommended when customers ask related questions, and gives a visibility score. This is the measuring side.

The second is a GEO specialist that produces content. It manages topics around user questions, writes drafts based on brand facts and flagship products, files content under the matching topic and can also generate WeChat official account drafts. Publishing is carried out by a publishing machine executing the GEO plan.

To be clear about what it cannot do: search indexing and citation by online AI require separate verification, and monitoring results must not be treated as proof of indexing or citation. Scheduled monitoring observes brand perception under offline models. Actual publishing still requires publishing authorisation on your own devices; the system will not bypass platform rules for you.

The judgment to make now

Progress such as GPT‑6 Cyber will not directly change footfall in your shop any time soon. But it is a reminder: AI is turning from a tool into an executor. As execution becomes cheaper, what becomes scarce is being chosen.

Whether your brand is mentioned in AI answers, and whether it is recommended, is something that can be monitored, optimised and needs time to build. Knowing earlier where you stand for visibility across the six engines tells you earlier where to put your effort.

Whether to act now depends on whether your customers have already begun using AI to make purchasing decisions. If they have, this belongs in next year's plan.

閱讀繁體中文版 →