Artificial intelligence helps businesses automate processes and carry out many other tasks. However, as its capabilities grow, new security risks are emerging. AI systems not only detect vulnerabilities in networks and software, but under certain conditions can also exploit them independently. Several incidents that came to light in the summer of 2026 demonstrate just how far these capabilities have already progressed. During security tests carried out by OpenAI and Anthropic, AI models unexpectedly went beyond the boundaries of their test environments. This brings a new question to the fore: what happens when AI systems independently carry out security-critical actions that their developers did not anticipate?
Autonomous AI agents create new vulnerabilities
AI agents open up numerous opportunities for businesses, for example by planning tasks or running programmes independently. At the same time, this leads to increased security requirements.
If, for example, AI agents are granted access to internal databases, email systems or development environments, faulty configurations or unexpected actions can have serious consequences. Confidential information could be disclosed or systems inadvertently altered.
Added to this are so-called prompt injection attacks. In these attacks, cybercriminals attempt to manipulate AI systems via tampered websites, documents or other content. An AI agent could mistakenly interpret hidden instructions (“white text on a white background”) as part of its task and carry out undesirable actions. Furthermore, AI systems cannot always reliably distinguish between data and instructions. For businesses, this means that not only traditional IT systems, but also AI applications and their interfaces must be consistently secured.
Are cyberattacks becoming more dangerous because of AI?
The incidents described do not mean that every AI system is capable of bypassing all security mechanisms unchecked. Nevertheless, they illustrate just how rapidly their capabilities are evolving. Cybercriminals are already using artificial intelligence to automatically generate personalised phishing messages. AI is also being used to gather and analyse information about potential victims or to prepare attacks. In future, autonomous AI agents could automate further steps in the process, thereby increasing the speed and scalability of cyberattacks.
This is putting increasing pressure on businesses to patch security vulnerabilities – such as the recently disclosed vulnerability in the PHP programming language – at an early stage and to detect suspicious activity as quickly as possible. The more automated attacks become, the less time businesses may have to respond effectively.
How businesses can improve their security
These new risks do not require an entirely new security strategy. Rather, tried-and-tested protective measures must be consistently extended to AI applications. Of particular importance are clearly defined and task-specific access rights for AI agents, secure and isolated test environments, and human approval for critical or irreversible actions. Organisations should also be able to track which tasks AI systems are performing and which data they are accessing. Another important measure for enhancing security is the implementation of two-factor authentication for logging into user accounts.
Continuous monitoring of the IT infrastructure is also becoming increasingly important. Modern security solutions utilise, amongst other things, AI and behaviour-based analytics to identify and stop suspicious activity at an early stage – for example, through Managed Extended Detection and Response (MXDR) or Extended Detection and Response (XDR). The interplay between modern technology, clear security processes and human expertise remains crucial.
Conclusion: AI needs clear safety limits
Recent incidents show that as the capabilities of artificial intelligence grow, so too do the demands on cyber security. Autonomous AI agents, in particular, pose a risk if they are granted far-reaching permissions and their actions are not adequately monitored.
It is therefore crucial for organisations to design AI systems securely from the outset, to restrict access rights and to detect unusual activity at an early stage. The problem is not artificial intelligence itself, but the scope granted to it.
FAQ: Autonomous AI and Cybersecurity
Can AI systems carry out cyberattacks on their own?
Under certain conditions, powerful AI agents can independently exploit vulnerabilities and gain access to external systems. Such capabilities were observed during tests this year. However, this does not mean that ordinary AI applications are automatically capable of carrying out cyberattacks on their own.
What is the difference between an AI agent and a chatbot?
Whilst chatbots mainly generate responses and content, AI agents can independently plan tasks, use tools and interact with other IT systems. This gives rise to additional security risks.
What is prompt injection?
In a prompt injection attack, attackers attempt to influence AI systems using manipulated content. The aim is to divert the AI from its intended task and trigger unwanted actions.
How can organisations secure AI agents?
Key measures include restricting access rights, monitoring actions carried out, using secure test environments and requiring human approval for critical operations. AI applications should also be integrated into the existing IT security framework.
