New tools, old bugs: risks related to AI

Relying too much on AI can be dangerous. At the beginning of June this year, a malicious program called Miasma took over more than 70 of Microsoft’s digital databases. A few days later, it also appeared in the program download system (PyPI) and began to infect so-called components, i.e., ready-made pieces of code. As Exorigo-Upos notes, the most important thing in this story is how the malicious program hides and maintains access to the system. Miasma saves itself in the settings of Claude Code, an AI tool for developers. This allows it to restart with each use of the tool and steal sensitive information such as passwords or passkeys.

In the past, malicious programs added themselves to the system settings of a computer to launch during startup. Now they are doing the same, but inside AI tool settings. However, users rarely check the settings of AI tools. And it is exactly there that important access data for company systems is often located today.

“The use of AI in organizations is becoming more and more common. However, conversations with our clients show that few of them understand the risks associated with the use of artificial intelligence, and even fewer know how to effectively defend themselves against them. The security measures we implement should be a direct result of the risk analysis carried out in the organization. However, the question is how to effectively conduct such an analysis if we do not understand the technology and do not see the risks it brings,” comments Michał Tomaszewski, CISO, Security Team Manager at Exorigo-Upos and Co-founder & Board Member of Knoxtera.

The most common risks associated with AI

Many organizations apply the principle of least privilege. Nevertheless, this principle is applied selectively. Local admin privileges aren’t restricted for some teams, often including developers. Companies assume that these are security-aware users, so the risk is low. However, if an organization uses AI agents to create code, the risk landscape changes dramatically. The permissions of the local administrator, even for development groups, create additional risks that the company was not previously exposed to.

“Claude Code tools run locally on the developer’s workstation and inherit user permissions. In such a model, we make one of the biggest mistakes: we give artificial intelligence broader privileges than it should have. This can lead to a situation where AI will have access to data that it shouldn’t, such as keys and credentials, or even the ability to perform unauthorized operations. In the worst case, it can lead to damaging a given environment by deleting files, changing the configuration, or running a script,” notes Michał Tomaszewski.

However, these threats are not the only issue. Many companies are convinced that the introduction of AI policies is sufficient security. Entrepreneurs believe that they are free from threats because employees are informed that they cannot upload specific types of data to the models, e.g., constituting a trade secret or containing personal data, or that they should use commercial licenses. However, simply using paid versions of the tools does not protect the organization from data leaks or loss of information. If the license is purchased by an employee and is personal, the company loses control over the settings for the indicated models, e.g., whether AI can use the data for learning. In addition, after an employee leaves, the organization loses what is in their private account, and the employee themselves may still have access to the company’s data even long after changing jobs.

“Many companies do not control what is input into AI. Simply having policies that say not to include customer data in the models may be insufficient. The simplicity and time savings of using AI are just too tempting. Therefore, if we as an organization do not completely block external models and, for example, do not provide our own tools, we should implement security systems designed to monitor, control, and block sensitive data. Otherwise, the greatest value of many organizations, i.e., their know-how, is not effectively protected. Therefore, when conducting a risk analysis, it is crucial to understand the technology and use proven methodologies. In addition, when implementing new tools or technologies, it is worth conducting a risk analysis again, because sometimes even a seemingly small change can have a significant impact on the level of risk,” explains Michał Tomaszewski.

Growing number of threats

According to the latest analyses, the number of detected vulnerabilities in AI systems is growing rapidly. According to the TrendAI™ State of AI Security report, more than 2,000 new security vulnerabilities were recorded in 2025 alone, which means a year-on-year increase of more than 30 percent. TrendAI™ experts also point out that artificial intelligence systems are becoming an increasingly frequent target of attacks. The number of detected vulnerabilities in the area of AI is growing faster than in the entire IT ecosystem. In 2025, such systems were already responsible for more than 4% of all reported vulnerabilities in software – this is the highest result in history.

In total, more than 6,000 security vulnerabilities related to artificial intelligence have been identified in recent years. Importantly, a significant proportion of them have a high or critical threat level, which means that they can lead to system takeovers, data breaches, or serious disruption of infrastructure. An increasing number of vulnerabilities also affect high-access AI agents and MCP servers, which allow models such as Claude or ChatGPT to securely connect to external tools in real time.

Don’t leave your business vulnerable to new threats. The development of AI is a huge opportunity, but it also brings new risks that require a modern approach to data protection. Are you wondering if your infrastructure and implemented policies are fully secure? Contact us. Our cybersecurity experts will be happy to answer your questions and help you conduct a comprehensive audit to identify vulnerabilities and secure your company’s know-how.