[AI Hacking] Security Competition Heats Up as AI-Powered Cyberattacks Grow More Sophisticated; OpenAI, Anthropic and Google Accelerate Market Push

AI-powered cyberattacks surge, with even autonomous AI agents deployed for hacking
AI also abused in PHP2Shell attacks and Binance hack, expanding security threats to businesses
OpenAI, Anthropic and Google roll out AI-based systems for detecting and fixing vulnerabilities

Cyberattacks using artificial intelligence (AI) are rapidly becoming more sophisticated. With AI assisting in △ vulnerability discovery △ attack code development △ internal network reconnaissance, complex intrusions can now be carried out at lower cost than before. Recent cases have even shown AI agents analyzing attack situations and selecting appropriate tools and commands. In response, major AI companies, including OpenAI, Anthropic and Google, are accelerating the race to defend systems by releasing security models and solutions specialized in detecting and fixing vulnerabilities.

AI Agents Turned into Hacking Tools

According to a report on the 3rd by U.S. cybersecurity publication Help Net Security (all dates hereafter are local time), global security company Kaspersky recently reported cyberattacks targeting educational institutions in several African and Latin American countries using the open-source AI-powered penetration testing tool CyberStrikeAI. According to Kaspersky, the attackers infiltrated the victim institutions' internal networks, collected system information for a day and then paused their activity for about two weeks. The attacks subsequently escalated when CyberStrikeAI's command-and-control (C2) agent became involved. Within two minutes of deployment, the AI agent had established persistent access to the compromised systems and completed initial reconnaissance. Within 3–4 days, it had compromised domain administrator accounts.

Analysis indicated that the agent used a large language model (LLM) on the attackers' command-and-control server to determine intrusion paths and generate commands. Under this setup, human attackers specify the target and objectives, the AI selects tools and commands appropriate to the situation, and a program installed on the victim's system executes them. The attackers then used the access they had secured to steal large volumes of credentials stored across multiple computers and exfiltrate databases. These final stages of the attack took less than an hour. Kaspersky did not definitively attribute the incident to a particular hacking group, but suggested that Chinese speakers may have been involved, citing Chinese-language comments in the scripts used and traces of activity on Chinese online services.

AI Also Used in PHP2Shell Attacks

Such cases of AI-powered hacking are emerging around the world. A prominent example is the cyberattacks carried out last month by suspected China-linked hackers exploiting PHP2Shell. PHP2Shell collectively refers to two vulnerabilities (CVE-2026-109618 and CVE-2026-109619) disclosed by the official PHP development team on the 21st of last month, which allow attackers to remotely execute malicious code on vulnerable servers without authentication. The attackers combined these flaws with Fragnesia, a privilege-escalation vulnerability in the Linux operating system, to attack firewalls made by global cybersecurity company Palo Alto Networks. An investigation by Israeli cybersecurity company Dream found that the attackers scanned approximately 19,000 internet-exposed firewalls before narrowing their targets to about 240. Of these, they were found to have taken control of 24 firewalls and one internal management server.

AI was used in this process to develop intrusion tools and identify attack targets. Dream researchers discovered an attack platform called AIScan on the attackers' systems that integrated Anthropic's AI coding tool Claude Code with Chinese AI models DeepSeek and Kimi. The attackers are believed to have used the tool to identify the types and software versions of target firewalls and write attack code tailored to specific devices. However, the researchers noted that AI did not automatically carry out the entire intrusion process. While they found indications that AI assisted the attackers in preparing attacks and developing tools, the respective roles of humans and AI in specific decisions and execution stages remain unclear.

Binance's Internal AI Becomes an Attack Path

AI-related tools were also confirmed to have been used in a recent hack at Binance, the world's largest cryptocurrency exchange. Binance's internal investigation found that the attackers spent several months targeting employees, beginning in June. They first induced an engineer responsible for cryptocurrency wallet development and operations to download a malware-infected software installer while installing an AI coding tool. Binance's security team detected suspicious activity at the time, replaced the employee's computer and changed the credentials, but information about the employee's work and colleagues had already been leaked. The attackers then sent a fake job offer to another software engineer in September and persuaded the engineer to install a malware-infected program on a work computer under the pretext of completing an interview assignment. The malware worked by modifying part of a legitimate development tool to communicate with an external server, and it went undetected by Binance's security systems at the time.

The attackers began exploiting Binance's internal AI tools in earnest on the 20th of last month. Using the engineer's previously infected computer as a foothold, they accessed the exchange's in-house AI tools and gained control of the computers of two employees responsible for cryptocurrency wallet development and operations. During this process, they discovered a token granting container deployment privileges stored on one employee's computer and used it to establish a means of deploying malicious programs to wallet-related systems. During the actual attack on the 29th of last month, they used the previously stolen credentials to deploy containers and execute malicious scripts, then abused Binance's internal wallet transaction signing system to transfer cryptocurrency to external wallets without going through normal withdrawal procedures.

OpenAI Sets Its Sights on Cybersecurity

As advances in AI technology make cyberattacks increasingly sophisticated, the view that AI will also serve as a key means of countering these risks is gaining traction in the market. Major AI developers have indeed identified cybersecurity as their next growth engine and are accelerating efforts to build out enterprise solutions. OpenAI, for example, recently hired an executive from Wiz, the Israeli security company acquired by Google, to strengthen sales to enterprise customers. In an interview with Bloomberg TV at an international AI summit in Dublin, Ireland, on the 14th of last month, OpenAI Chief Financial Officer (CFO) Sarah Friar explicitly said that cybersecurity would be a major driver of the company's future growth.

The core technology underpinning OpenAI's push into cybersecurity is Daybreak, an AI-powered system for detecting and fixing security vulnerabilities. Unveiled in June, Daybreak analyzes companies' software code and systems to identify potential vulnerabilities, verifies whether they can actually be exploited, and supports the generation and testing of code fixes. In August, OpenAI also introduced its latest security-focused model, GPT-5.6-Cyber, and refined its framework for assigning access permissions according to users' qualifications and work purposes.

Table 1. Major AI Companies' Expansion into Cybersecurity

Anthropic and Google Also Enter the Market

In February, Anthropic unveiled Claude Code Security, which integrates security capabilities into its software development tool Claude Code. The solution analyzes a company's entire source code to identify vulnerabilities that conventional static analysis tools can easily miss, assesses their severity and exploitability, and proposes specific fixes. Anthropic further expanded the business in April with the release of its high-performance AI model Claude Mythos Preview, while also launching Project Glasswing with Amazon Web Services (AWS), Apple, Microsoft, Google, JPMorgan Chase, CrowdStrike and others. Project Glasswing is a joint security program that helps major technology and financial companies use Mythos to proactively identify and fix security vulnerabilities in their software and critical systems. Participating organizations reportedly discovered more than 10,000 high-severity or critical security vulnerabilities in just over a month after the project's launch.

Google also launched its Fairwind program in September for government agencies, major companies and cybersecurity partners. Similar in nature to Project Glasswing, Fairwind provides approved organizations with limited access to Google's advanced Gemini-family models and related technologies. Participating organizations can use AI to automatically search for potential software vulnerabilities and develop fixes. Multiple AI models, including Google's Gemini 4 Argon, were reportedly also used in Wiz's recent work to discover critical PHP vulnerabilities, detect signs of exploitation and validate fixes.