Main points
- Google Threat Intelligence Group has discovered the first zero-day exploit designed entirely by artificial intelligence to bypass two-factor authentication in systems.
- Cybercriminals and state-sponsored hacking groups are actively using AI to create and disguise malware, as well as to analyze vulnerabilities in software.
Artificial intelligence has gone on the offensive: Google recorded the first case of creating an AI exploit / Collage 24 Channel / Unsplash
Google’s Threat Intelligence Group has released a report on a disturbing event that marks the beginning of a new era in cybersecurity. For the first time in history, a zero-day exploit designed entirely with the help of artificial intelligence has been discovered. It was reportedly preparing for a large-scale attack on digital networks.
Why is the emergence of AI exploits becoming a challenge for global security?
Attackers have begun to move from simple experiments to full-fledged use of generative models in their operations. The most high-profile case was the discovery of a zero-day exploit created by artificial intelligence at the behest of a criminal group for mass hacking of systems. This tool was implemented in the form of a Python script designed to bypass two-factor authentication (2FA) in a popular open-source system administration tool, Gizmochina reports.
Code analysis of this exploit confirmed its artificial origin. The script contained a large number of training comments, a hallucinatory CVSS vulnerability score, and had the perfect structure typical of large language models.
While frontier AI models are still struggling to understand the complex logic of enterprise authorization, they have an increasing ability to perform contextual reasoning, effectively reading the developer's intent,
– Google Threat Intelligence Group analysts commented on their report.
This allows AI to find logical errors that remain unnoticed by traditional automated scanners because the code appears functionally correct but contains strategic security miscalculations.
In addition to independent cybercriminals, state-run hacking groups are also showing an active interest in AI:
- In particular, groups linked to the PRC and DPRK use specialized datasets to find vulnerabilities, Tom's Hardware noted. For example, the UNC2814 group forced the Gemini model to act as a senior security auditor or C/C++ binary expert to analyze the firmware of TP-Link devices.
- The APT45 group sent thousands of repeated requests to automate the analysis of known vulnerabilities (CVE) and the verification of attack prototypes, which would be impossible to do manually at such a scale.
- Russian hackers are also not left out, using AI to create malware that can evade detection. The CANFAIL and LONGSTREAM virus families contain large amounts of AI-generated “junk” code that acts as a smokescreen to mask the real malicious functionality.
The explanatory nature of the comments surrounding the decoy code indicates that the attacker asked the model to intentionally create a large amount of inert code for obfuscation,
– reported in the Google Cloud Blog report.
A separate threat is autonomous malware, such as PROMPTSPY for Android. It uses the Gemini API to independently navigate the smartphone interface, interpret user actions, and execute commands without direct human intervention. PROMPTSPY is even capable of intercepting biometric data and replaying screen unlock gestures to regain access to the device.
AI itself is also becoming a target for hackers
At the same time, AI is becoming not only a tool for attack, but also a target. In 2026, attacks on the supply chain of AI components increased. The group “TeamPCP” (also known as UNC6780) hacked popular GitHub repositories such as LiteLLM to introduce credential thieves and steal access keys to AWS cloud services. This poses risks for organizations that integrate AI models into their internal processes without verifying the security of third-party libraries.
How to resist this?
To combat these threats, Google is implementing AI-powered defenses. The Big Sleep agent helps find unknown vulnerabilities in software before hackers exploit them, and the CodeMender tool automatically suggests fixes for critical code errors.
Thus, artificial intelligence becomes the main battlefield, where the speed and analytical capabilities of models determine who wins the next round of digital confrontation.
You may also be interested in: Cyberspace in the Age of AI – Why Artificial Intelligence is Turning Our Security into a Nightmare
Artificial intelligence has dramatically changed cybersecurity, and in two ways. On the one hand, companies have received new tools for finding errors in code, analyzing attacks, and protecting systems. On the other hand, the same technologies have begun to be actively used by hackers, state cyber groups, and cybercriminals, as 24 Kanal has noticed. This is why experts are increasingly saying that generative AI has become one of the biggest catalysts for modern cyberattacks.
The main problem is scalability. Previously, creating a convincing phishing email, writing malicious code, or finding vulnerabilities required either experienced specialists or a lot of time. Now, a significant part of this work can be done by AI. It allows you to automate attacks, accelerate target reconnaissance, and generate code in literally seconds.
Personalized hacking
This has had a particularly strong impact on phishing and social engineering. As 24 Kanal reminds, large language models can create well-written messages without the typical errors that previously gave away scammers. They also analyze information about the victim from open sources and form personalized messages. As a result, phishing emails increasingly resemble real corporate correspondence.
Researchers also warn about using AI to automate massive social engineering campaigns, according to a study published on arXiv.
In addition to text, AI is actively used to create deepfakes – fake voices and videos. There have already been cases where criminals imitated the voices of company executives to convince employees to transfer money or provide access to systems. Such tools significantly increase the effectiveness of attacks, especially in a corporate environment.
AI writes viruses
Another problem is the generation of malware. Modern models can help in writing viruses, backdoors, protection bypass scripts and code hiding tools. Although most large AI services have limitations, hackers regularly try to bypass them through so-called jailbreak methods.
In 2025-2026, researchers already described cases where models helped create entire phishing campaigns even for users without serious technical skills, according to another study on arXiv.
AI looks for vulnerabilities
A separate direction is the search for vulnerabilities. This is where AI has become especially dangerous. Automated systems can notice gaps that a person would never pay attention to. This allows hackers to literally comb the Internet in search of victims while the criminals themselves drink coffee.
AI developers are trying to counteract
Against this backdrop, technology companies have begun to actively develop “defensive” AI. One of the main ideas is to use models to automatically find errors in code faster than hackers can. That is why specialized systems for analyzing software security have recently appeared.
One of the most high-profile examples was the collaboration between Mozilla and Anthropic. Initially, the Claude Opus 4.6 model analyzed approximately 6,000 Firefox files over two weeks and found 22 vulnerabilities, 14 of which were classified as critical or high-risk. Mozilla confirmed the authenticity of the problems and fixed them in Firefox 148, TechSpot writes.
Later, the situation became even more serious. Mozilla used the experimental model Anthropic Mythos Preview to analyze the entire code base of Firefox 150. According to the company, the system found 271 vulnerabilities, most of which were real, not false positives. Among the problems found were bugs in WebAssembly, IndexedDB, DNS processing, process isolation mechanisms, and even in components where errors had remained undetected for over 15 years.
Mozilla said that the AI didn't just work as a “chatbot.” They created an entire agent pipeline for it: the model accessed testing systems, analyzed the code, formed hypotheses about errors, automatically created test scripts, and verified the results. This is what allowed them to drastically reduce the number of false positives.
Other companies are also developing similar systems. Google, OpenAI, Anthropic and major cybersecurity firms are investing in models that can automatically analyze code, find logical errors and detect dangerous configurations. At the same time, the researchers themselves acknowledge that a new arms race is emerging: the same tools can be used by hackers, writes Axios.