Artificial intelligence is no longer a marginal tool in cyberattacks. A report published by the Google Threat Intelligence Group in May 2026 said that generative artificial intelligence is being integrated into attackers’ workflows on an industrial scale, and that its use is now appearing in nearly every stage of an attack. According to the article published by ITmedia AI Plus, the group observed for the first time a threat actor attempting to use an exploit for a zero-day vulnerability developed with the assistance of artificial intelligence.
The vulnerability was used in an attempt to target an open-source web management tool and led to the bypassing of two-factor authentication. More importantly, it did not result from a conventional implementation error such as memory corruption, but from an advanced logical flaw in the application’s design. This indicates that large language models capable of reading context and reasoning may help uncover logical errors that traditional vulnerability-scanning tools have been unable to notice.
The Problem Is Not Artificial Intelligence Alone, but Its Ease of Access
The article says that the observed attacks relied primarily on commercial accounts for closed models that had been compromised or misused. However, the spread of open-weight models is changing the equation from a different angle: after a model’s weights are downloaded, its use becomes less subject to monitoring by the developer, and it becomes difficult for the publisher to withdraw it or impose subsequent controls on it.
Daisuke Fuji, who is responsible for cybersecurity operations at Accenture Japan, explained that the fundamental difference between closed and open-weight models concerns the developer’s ability to control their use. Closed models typically combine safety controls, account logging, terms of use, and activity monitoring. Open-weight models, by contrast, can be modified or retrained to weaken their tendency to reject harmful behavior, including producing versions described as “uncensored.”
Attackers do not need to possess the best model available. According to Fuji, even models that perform less well than Anthropic’s and OpenAI’s closed models can find a large number of vulnerabilities. Discovering deep vulnerabilities that have gone undetected for decades may require a more advanced model, but criminals seeking money generally prefer targets that can be attacked quickly and at low cost.
What Changes in Practice for Organizations?
These developments do not mean that attacks will immediately become fully autonomous operations run by machines. Fuji believes that the dominant use is still to assist humans, because operating advanced models at scale requires graphics processing units and server resources that are limited and expensive. The most realistic use at present is to accelerate reconnaissance: identifying potential websites, scanning a large number of targets, and finding systems that can be attacked quickly.
But lowering the cost of reconnaissance may expand the list of targets. Systems that once appeared economically unviable may become viable targets if information about them can be gathered automatically. This includes the financial and cryptocurrency sectors, as well as artificial intelligence companies themselves. The GTIG report indicates that artificial intelligence has simultaneously become an advanced driver of attacks and a high-value target. Attacks have been observed against the artificial intelligence software supply chain, including the insertion of malicious code into skill packages designed for artificial intelligence agents and the compromise of gateways connecting multiple large language models.
Defense Begins by Assuming a Breach Will Succeed
The most important practical conclusion in the article is that preventing every attack is no longer a realistic foundation for building a protection plan. Fuji recommends taking measures that reduce damage after a breach, such as establishing business continuity plans, retaining data backups, and dividing networks into smaller zones through microsegmentation. These measures do not prevent every breach, but they aim to limit the attacker’s spread and accelerate the restoration of operations—principles that are well known in dealing with ransomware.
Reducing publicly exposed information also becomes more important. A model can gather small signals, such as technologies disclosed at public events or details in the code of web pages, and then assemble hypotheses about an organization’s environment. Fuji presents defensive deception—that is, providing false information to mislead an attacker’s reasoning—as one possible option.
Managing Patch Debt and Using Artificial Intelligence with Caution
The article also highlights the problem of “patch debt”: the accumulation of vulnerabilities left unaddressed because testing updates, preparing procedures, and verifying them after deployment are difficult. Fuji says that Japanese companies may be particularly affected by the complexity of their systems and the time required for the processes before and after installing patches. Among the trends he identified are reducing the number of systems that need patching by moving to the cloud and using artificial intelligence to accelerate testing and verification.
However, using artificial intelligence for defense carries a corresponding risk. If an organization sends its code or environment data to a provider and that provider is then attacked, the protection tool may become a source of information for the attacker. Organizations should therefore verify where data is stored and whether it will be used for training before building an artificial-intelligence-based vulnerability-scanning environment. The article notes that interest in sovereign artificial intelligence, which keeps the tool under the organization’s control, is growing among Accenture’s clients.
The editorial reading from certi.news is that the actual change is not the emergence of fully autonomous attacks, but the reduced cost of searching for exploitable targets and vulnerabilities. This increases the value of operational resilience and reducing the exposed information surface, alongside proactive scanning, penetration testing, and red-team exercises. The source emphasizes that defenders have an important advantage: their broader knowledge of their environments. However, the extent to which they benefit from this advantage will depend on data quality, information-sharing controls, and organizations’ ability to address what they discover before the attacker gets there first.