PHP exposed two high-severity vulnerabilities, hackers attempted millions of attacks, industry says AI is accelerating the imbalance between offense and defense

Download the client

+1 PHP security vulnerabilities shake the world: hackers make millions of attack attempts, the pace of offense and defense changes in the AI era. The Paper reporter Zhang Jing 2026-10-23 20:35 Source: The Paper ∙ Future 2% > Listen to full text Font size

For people working in cybersecurity, the past two days have been a dark moment. PHP, the programming language that underpins back-end development for three-quarters of the world's websites, was found to contain two severe security vulnerabilities. After the vulnerabilities were made public, millions of hacking attempts occurred, forcing multiple corporate and government websites around the world to be temporarily shut down until the vulnerabilities were fixed.

Relevant Chinese authorities have also taken note of this incident. On October 22, the Cybersecurity Threat and Vulnerability Information Sharing Platform of the Ministry of Industry and Information Technology issued a notice recommending that relevant units and users immediately carry out comprehensive inspections, promptly upgrade to the latest secure version, or adopt temporary protective measures such as taking services offline, disabling the session upload progress function, network isolation, and network access restrictions.

What exactly are the two exposed PHP vulnerabilities? How harmful are they? What new signals does this incident reveal about cybersecurity governance in the AI era? Multiple industry insiders told The Paper's Science and Technology desk that with the improvement of AI capabilities, AI is accelerating vulnerability discovery; high-risk vulnerabilities can be dug out in dozens of seconds, but the speed of vulnerability discovery is far faster than remediation, leading to an imbalance between offense and defense. In the cybersecurity field, defense is more difficult than attack, and remediation is more difficult than discovery.

Hackers make millions of attack attempts, global emergency patching

PHP is a computing language like HTML and Python. Wang Liejun, a security expert at Qi An Xin Artificial Intelligence Company, told The Paper's Science and Technology desk, "PHP is like the foundation of many websites."

The vulnerabilities in PHP had been lurking in the software for years, until two severe security vulnerabilities (code-named PHP2Shell) were published on October 21, with vulnerability numbers CVE-2026-109618 and CVE-2026-109619. These vulnerabilities can cause unauthenticated remote attackers to trigger memory corruption by sending specially crafted POST requests and ultimately achieve remote code execution, thereby gaining server control permissions.

"Attackers do not need an account or password; they only need to send a carefully constructed web request to execute arbitrary commands on the server. It is equivalent to someone entering your home without a key and being able to rummage through everything at will. Qi An Xin CERT has successfully reproduced this vulnerability and rated it as high-risk," Wang Liejun said.

Pan Xudong, associate researcher at Fudan University and full-time mentor at Shanghai Innovation Institute, said that for enterprises, attackers can directly obtain control of the host carrying the website through the PHP vulnerability, and then perform arbitrary operations on that machine. What enterprises are truly worried about is not just the compromise of a single device, but that attackers may use it as a springboard to further penetrate other devices in the enterprise's internal network.

Although the vulnerability discovered this time is not in the latest version of PHP but in an old version, so in theory not all systems will be affected. But Pan Xudong said that having many vulnerabilities in old software is itself a common problem in the cybersecurity field, especially since some long-established enterprises often do not update software frequently, resulting in a large exposure surface.

Taking banks and other financial institutions as an example, TJ Marlin, CEO of enterprise AI security company Guardrail Technologies, once said that the technology architectures used by these institutions combine the most advanced tools with software from decades ago, which may expose a large number of vulnerabilities. Because many banks use the same vendors and the same solutions, Marlin said this could become a vulnerability multiplier, making any AI-driven attack "potentially catastrophic in scale."

Because before the patch was made public, vulnerability verification code and technical details were leaked before remediation measures, leading to frequent real-world attacks.

Qi An Xin Xlab's large-network threat analysis platform captured nearly a thousand source IPs from multiple countries around the world attempting to use the PHP remote code execution vulnerability to launch tens of thousands of attacks externally. Captured malicious behaviors include backdoor implantation, malicious command execution, mining program deployment, remote control connections, malicious payload downloads, and server credential/data stealers.

Cloud service companies Cloudflare and Amazon said that as enterprises and governments deployed patches and preventively suspended websites, hacking attempts surged to millions. Cloudflare CEO Matthew Prince posted, "Ten hours before the public announcement, we saw thousands of attempts to exploit this vulnerability across the entire network."

Due to the severity of the vulnerabilities, websites in multiple countries were urgently taken offline on a large scale. From Tokyo to Helsinki, numerous websites for banking, aviation, rail, tax, and government services were urgently shut down. Due to concerns about rising hacking risks, South Korea raised the cybersecurity alert level for the private sector and activated emergency response measures. Felix Kuhlenkamp, security head of the German digital industry association Bitkom, said that many large operators have completed updates, but for complex IT systems and small organizations, it may take longer.

Why did the vulnerability lurk for years and only be discovered now? Wang Liejun said that such vulnerabilities are deeply hidden and are triggered only under very special conditions, such as the moment when server memory happens to be exhausted, or the instant when the previous request has just ended and the next request is not yet ready. Normal business traffic almost never reaches these paths, and manual auditing also finds it difficult to exhaustively enumerate all combined scenarios. PHP has a huge amount of code and a heavy historical burden, and memory management logic is scattered across the engine, server interfaces, and various extensions, making it difficult for traditional auditing tools to track state changes across requests. These blind spots are being covered by AI's batch analysis capabilities, and the pace of discovery of similar kernel defects has clearly accelerated.

This PHP vulnerability was discovered by cybersecurity company Calif using GPT‑5.6 Sol, and Google's cybersecurity company Wiz also independently discovered and reported these two PHP vulnerabilities.

Pan Xudong told The Paper that before ChatGPT, hackers and security companies were already using automated vulnerability discovery methods, continuously testing, crafting tricky inputs, and observing whether systems showed anomalies to find problems. But limited by resources and targeting capability, hackers found it hard to conduct long-term automated searches and to precisely lock onto specific vulnerabilities. Now, however, AI agents can work around the clock and possess extensive cybersecurity knowledge, enabling them to undertake this kind of continuous, targeted vulnerability mining. In particular, since Anthropic released the Mythos model in April this year, the cyberattack capabilities of large models have markedly increased.

OpenAI expects AI to play an increasingly large role in helping defenders discover vulnerabilities, monitor systems, and respond to threats. "Over the past two months, we have led the formation of an alliance of maintainers, researchers, and security service providers to jointly address severe vulnerabilities in PHP... Given the severity of these vulnerabilities and the widespread use of PHP, we believe this is one of the most impactful security issues in recent years," OpenAI said. "Although these vulnerabilities are significant, they are only part of the multiple major vulnerabilities we have already discovered and are working to fix. We expect similar discoveries to increase substantially in the coming months."

Wang Liejun said that in the AI era, vulnerabilities are becoming an important topic in security governance. Vulnerability discovery has become assembly-line production, and the pace of offense and defense has been compressed from days to hours or even minutes. Take a large model like Mythos as an example: it can already independently complete vulnerability probing, weaponized code generation, and full-chain penetration. In the past, it took a senior engineer an average of 3-7 days to find a high-risk vulnerability; now it has been compressed to minutes or even tens of seconds. Just one month after Anthropic's "Glasswing Project" went online, Mythos helped partners uncover more than 10,000 vulnerabilities; Anthropic itself scanned open-source projects and found 23,000 vulnerabilities in 1,000 projects, with one week's output exceeding the global monthly average total of CVEs (Common Vulnerabilities and Exposures). This means that kernel-level vulnerabilities that had lain dormant for years are being unearthed in batches by AI. This is no longer a question of "whether" but of how fast.

Cai Yong, co-founder of Network Sharp Knife, told The Paper that regardless of whether AI is used, a large number of vulnerabilities are exposed every year. In the past two years, large model capabilities have matured, objectively accelerating vulnerability discovery, and defense work has shifted from checking logs daily in the past to monitoring logs in real time and handling security incidents. AI has also further highlighted the problem of vulnerability disclosure. Take the PHP vulnerability as an example: before this patch was made public, related exploits had already leaked through unofficial channels, making timely patching difficult and emergency response more challenging for security practitioners. "In the future, many vulnerabilities may not be discovered manually but through AI, and may not go through public processes but instead disclose vulnerability information in their own way. This will lead many people to target vulnerabilities for secondary exploitation, carrying out destructive acts or seeking profit."

Vulnerability discovery is far faster than remediation, leading to an imbalance between offense and defense

Stronger model capabilities have lowered the cost and skill threshold for discovering and exploiting vulnerabilities, but defensive capabilities have not kept up. Pan Xudong said that models do not yet have remediation capabilities at the same level as vulnerability discovery, creating the current offense-defense imbalance. In cybersecurity, defense has always been harder than attack, and fixing vulnerabilities is harder than finding them. "After fixing, you also need to prove that the fix itself is secure and does not introduce other problems. For attackers, as long as the vulnerability actually exists, it can be exploited."

Wang Liejun said the difficulty in defending against this PHP vulnerability lies in its broad impact: about 26 million network assets across the internet may fall within the potential impact scope of the vulnerability. "Patching" can cover only a limited scope, and a large number of systems are still using old PHP versions that the official maintainers stopped supporting long ago, with no patches available at all. In addition, defenders are left with very little time. The window from vulnerability disclosure to exploitation has been extremely compressed, possibly to only a few hours, and the old rhythm of defenders waiting for patches and then upgrading has completely failed. The attack surface is sinking from the application layer to the underlying layer. When a low-level component has a problem, all applications above it suffer as well. This PHP vulnerability is a typical example.

On governance, Wang Liejun believes in "using AI to defend against AI." At the enterprise level, automated penetration testing, AI-assisted asset inventory, and vulnerability verification should be embedded into daily security operations, making vulnerability discovery a routine task rather than an emergency action. At the regulatory level, he recommends promoting a coordinated vulnerability disclosure mechanism to avoid situations such as details leaking in advance or mirror sites releasing patches first, which compress the defense window. At the open-source community level, a normalized mechanism for AI-assisted code auditing needs to be established, with priority reinforcement of kernel-level state management logic.

Pan Xudong said that one focus of cybersecurity capabilities in the next stage is to improve AI's ability to fix vulnerabilities; otherwise, it will be difficult to truly protect future enterprises. Second, attention must be paid to the risk of AI losing control. An OpenAI agent chose to cheat in order to obtain higher scores in an evaluation task and used cyberattack capabilities to intrude into Hugging Face and steal answers. This is a loss-of-control problem. What people need to care about is whether AI will "do whatever it takes" to achieve its goals when under pressure or seeking to cheat, so the risk of AI losing control needs to be better understood and prevented. "This incident also shows that software still needs to be updated in a timely manner. This is not an AI issue; vulnerabilities that were not discovered in the past will eventually be discovered."

Cai Yong said cybersecurity risks always exist. After a common development framework vulnerability is disclosed, scanning traces may appear worldwide within 15 to 30 minutes. From the current situation, although many organizations continue to purchase security devices and products, they generally lack professional operational capabilities. Devices generate a large number of alerts, but it is difficult to determine whether they are real attacks or scanning probes, ultimately leading to alert fatigue, while hacker attacks are often mixed in with vulnerability probing and attack behavior.

"Security operations require professionals to upgrade detection rules in real time and identify anomalous behavior; security devices cannot guarantee that intrusions will never happen 100 percent," Cai Yong said. Whether or not vulnerabilities such as the PHP one emerge, security operations capabilities should be continuously improved in daily work, rather than only responding urgently when a vulnerability appears and relaxing when none exists. Hacker attacks happen all the time, and it is not enough to pay attention to security only when a development framework or component is exposed with a vulnerability. "No matter what vulnerability there is, we need to focus more on whether we can discover and handle attack behavior at the first opportunity." Executive Editor: Huan Yanhong Image Editor: Jiang Lidong Proofreader: Zhang Liangliang Pengpai News tip line: 021-962866 Pengpai News, no unauthorized reproduction +1 Favorite I want to report #PHP security vulnerability #hacker attack View more