DriveLock Blog | IT Sicherheit und Cyber Security

How Does AI Pose a Threat to IT Security?

Written by DriveLock | Jul 22, 2026 8:15:00 AM

The rapid adoption of artificial intelligence is fundamentally changing companies’ defense mechanisms. While these technologies promise enormous efficiency gains, they also create entirely new entry points for cybercriminals.

CONTENTS
  1. WHAT IS AI SECURITY?
  2. MAIN RISKS OF AI SECURITY
  3. 5 EXAMPLES OF HOW AI THREATENS CYBERSECURITY
  4. HOW CAN COMPANIES IMPROVE THEIR CYBER DEFENSES?
  5. CONCLUSION AND SUMMARY


Critical infrastructure, manufacturing, and the healthcare sector, in particular, are increasingly becoming targets of automated attacks. It is therefore essential for IT managers in the DACH region to precisely analyze this evolving threat landscape. Traditional security approaches are often no longer sufficient to effectively counter the new, dynamic malware. This article highlights the most serious risks and presents practical solutions for your digital defense strategy.

A. What Is AI Security?


The term “AI Security” describes a dual concept: On the one hand, it encompasses the protection of AI systems against manipulation, data theft, and misuse. On the other hand, it refers to the use of artificial intelligence to defend against digital attacks. A central and modern component of this framework is what is known as agentic AI (agent-based AI).

Unlike passive language models, which merely generate text, AI agents are characterized by a high degree of autonomy. They are capable of independently analyzing complex objectives, planning intermediate steps, utilizing external tools or application programming interfaces (APIs), and making decisions without human intervention.

AI in Cybersecurity

In the field of cybersecurity, these autonomous systems act like digital security analysts operating at the speed of light. On the defensive side, they continuously monitor networks, interpret error messages, and autonomously isolate infected servers before any damage occurs. On the attacker side, however, malicious AI agents carry out fully automated hacking attacks. They dynamically adapt their strategy to the security architecture they encounter, bypass firewalls, and find vulnerabilities without human control.

B. Key Risks of AI Security


The implementation of autonomous systems into corporate IT entails specific risks that go far beyond traditional software errors. If AI agents are granted extensive access rights, unforeseen interactions with corporate data can create catastrophic security vulnerabilities. The following core risks currently threaten modern IT infrastructures:

  • Indirect Prompt Injection: In this scenario, attackers manipulate external data sources—such as a received email or a PDF document—with hidden control commands. As soon as the AI agents read this manipulated file, the external commands override the internal security rules. The agent then unintentionally performs malicious actions, such as disclosing internal passwords.
  • Privilege Escalation: To operate autonomously, digital assistants often require extensive system privileges. If these privileges are not strictly limited, compromised AI agents can independently expand their authority or combine legitimate tools in such a way as to gain access to sensitive HR or patient data.
  • Endless Logic Loops (Agentic Looping): Attackers can intentionally flood systems with unsolvable tasks or contradictory commands. This causes the affected agent to enter an infinite loop, consuming massive amounts of computing power and incurring immense cloud charges. In the worst-case scenario, this leads to an artificial system outage (Denial of Service).
  • Data Leakage via Shadow AI: Employees frequently use public, unauthorized AI tools in their daily work to review program code or draft reports. In the process, sensitive company secrets, patient data, or production plans flow unnoticed to external servers, resulting in a complete loss of control over where the data ends up.

 

C. 5 Examples of How AI Jeopardizes Cybersecurity


Recent real-world incidents demonstrate that these theoretical risks have long since become economic reality. Maliciously manipulated or carelessly integrated systems have already caused significant financial and structural damage. The following five case studies illustrate how AI agents and generative algorithms have successfully breached existing security barriers:

1. The $25.6 Million Fraud via Deepfake Video Conference (Arup)

The Vulnerability: Social engineering, amplified by real-time generative AI.

The incident: In early 2024, an employee in the finance department of the British engineering firm Arup was targeted by fraudsters using deepfake technology. The employee was invited to a video conference to discuss a confidential transaction. All other participants in the conference—including the company’s CFO and several colleagues—were deepfakes generated in real time by AI.

The damage: Because the employee relied on visual and auditory confirmation from people he believed he knew, he bypassed standard protocols and authorized several wire transfers totaling $25.6 million. This case remains a landmark example of how AI can completely undermine standard identity verification and trust structures within a company.

2. The Security Vulnerability in Slack AI (Indirect Prompt Injection)

The Vulnerability: Indirect prompt injection via LLM integrations.

The Incident: Cybersecurity researchers uncovered a massive vulnerability in the way Slack AI (the LLM assistant integrated into the platform) processed data. Since the AI was designed to gather and summarize information across channels, attackers realized they could inject malicious, invisible commands into a public document or message. When an unsuspecting user asked Slack AI to summarize that specific thread or channel, the hidden malicious commands hijacked the AI’s logic.

The result: The AI assistant was tricked into ignoring its built-in security measures. It secretly extracted sensitive data from the user’s private Slack channels and forwarded the information to an external server controlled by the attacker—completely undermining the internal confidentiality and security controls of companies that rely on Slack for secure collaboration.

3. The “EchoLeak” Vulnerability in Microsoft 365 Copilot

The Vulnerability: Data exfiltration via automated email processing.

The incident: Security researchers identified a zero-click vulnerability called EchoLeak in Microsoft 365 Copilot. An attacker could send a specially crafted email containing hidden, malicious AI instructions to a corporate inbox. As the automated Copilot system indexed and analyzed the email in the background to build its context database, the hidden AI instructions took control of the workflow.

The result: Without a human user ever clicking a link or opening an attachment, the AI interpreted the malicious instructions as valid prompts. It was forced to secretly collect sensitive corporate data (such as internal emails, calendars, or documents) and transmit it to an external third-party domain, thereby bypassing conventional email security filters and perimeter defenses.

4. Automated Cyberspying Against Government Networks

The Vulnerability: AI-powered vulnerability scanning and exploit generation.

The incident: Reports from the threat intelligence community pointed to a serious security breach in which a hacker managed to bypass the security measures of Anthropic’s LLM “Claude” to orchestrate a massive cyberattack on the Mexican government’s infrastructure.

The Damage: By tricking the AI into “cracking” its own security protocols, the attacker used the model to autonomously scan the network infrastructure, identify vulnerabilities, and generate clean exploit code. This AI-accelerated attack enabled the attacker to steal 150 gigabytes of sensitive tax, voter, and login data, proving that commercial AI can be weaponized to drastically lower the technical barriers to cyberespionage with far-reaching consequences.

5. “Shadow AI” and the Misuse of Corporate Access Credentials

The Vulnerability: Unregulated use of consumer AI in enterprises (“Shadow AI”).

The incident: According to data from IBM’s X-Force Threat Intelligence, over 300,000 sets of ChatGPT user credentials were discovered being traded on the dark web after being harvested by infostealer malware. Employees at hundreds of large companies had regularly copied and pasted proprietary code, intellectual property, and internal system logs into browser-based AI chatbots to make their daily work easier.

The damage: Because employees bypassed the company’s official procurement channels (leaving a massive “shadow AI” footprint in the process), security teams had no visibility into what data was leaving the network. When hackers stole browser cookies and login credentials for these AI accounts, they gained immediate, unauthorized access to months’ worth of proprietary data and internal API endpoints that had been cached in the AI command history.

D. How Can Companies Improve Their Cyber Defenses?


To protect themselves against the threat scenarios described, organizations must urgently adapt their defense strategies. A static firewall is no longer sufficient against dynamically operating, malicious AI agents. By implementing the following measures, companies in industry, healthcare, and critical infrastructure can sustainably strengthen their resilience:

  1. Implementation of isolated runtime environments (reasoning sandboxes): Autonomous systems should only be allowed to execute critical commands or data analyses within isolated virtual environments. This prevents a compromised AI from causing direct damage to the main network.
  2. Establishing the principle of least privilege: Digital assistants must never be granted blanket administrator rights. Any action with far-reaching consequences—such as deleting databases or sending data externally—must be subject to manual approval by a human (human-in-the-loop).
  3. Implementation of real-time behavioral monitoring: Since traditional signature scanners fail to detect AI-generated malicious code, companies must rely on continuous behavioral analysis. If an internal data stream or the behavior of an IT system suddenly deviates from the normal pattern, automated emergency shutdowns (kill switches) must be triggered.
  4. Regular Conduct of AI-Specific Stress Tests (Red Teaming): IT departments should systematically test their defenses using simulated attacks by malicious AI agents. This is the only way to detect logical vulnerabilities in their own AI models and connected interfaces at an early stage.
  5. Binding guidelines for the use of external tools: Clear compliance requirements and technical restrictions must prevent employees from entering business-critical information or patient data into public chatbots.

E. Conclusion and Summary


Advances in the field of artificial intelligence present entirely new challenges for corporate security. Autonomous AI agents are now capable of independently analyzing defense mechanisms and executing targeted attacks in milliseconds. As the real-world examples show, the financial and data protection consequences range from multi-million-dollar losses to the leakage of highly sensitive trade secrets.

Sectors such as healthcare and critical infrastructure—where system failures can threaten human lives—are particularly at risk. To thrive in this dynamic environment, IT leaders in the DACH region must rethink their approach. Protection against manipulated data streams and strict monitoring of all AI operations must henceforth be top priorities. Only through a combination of human oversight and modern behavioral analysis can digital resilience be ensured in the long term.

Since autonomous AI agents operate directly on your endpoints, this creates entirely new security risks for your company. DriveLock puts an end to this digital blind flight and immediately makes local AI tools—as well as their access to endpoints—transparent and controllable. Using proven concepts such as application and behavior control, you can effortlessly establish clear guidelines without hindering the productive use of innovative technologies.

This allows you to ensure long-term adherence to your governance and compliance requirements while simultaneously reducing the manual effort required for your IT operations. Take the first step toward controlled AI security today and test DriveLock’s advanced endpoint solutions for Agentic AI!