Hackers Aren’t Just Attacking Your Computers. They’re Manipulating Your AI.

Click here to view/listen to our blogcast.

For years, cybersecurity has focused on preventing hackers from breaking into computers, stealing information, and installing malicious software. Artificial intelligence is now introducing another challenge.

What happens when hackers don’t just attack your computer, but also try to convince the AI protecting it that nothing is wrong?

Two recent cybersecurity reports illustrate how criminals are exploring this new territory. One describes malware designed to deceive AI security tools. The other reveals how attackers used an innocent-looking poem to help control thousands of infected servers.

Together, they raise an unsettling possibility: What happens when attackers combine these techniques?

When Malware Tries to Fool AI

Researchers at Cisco Talos recently demonstrated how malicious software can contain instructions intended to mislead AI systems analyzing it. Instead of simply hiding its malicious behavior, the software might tell the AI that it’s harmless or doesn’t require further investigation.

This technique, called prompt injection, attempts to make an AI system follow instructions embedded in information it is supposed to examine. Some techniques succeeded in influencing the AI’s conclusions in approximately 35% of test runs. Although far from guaranteed, even occasional success could be valuable to criminals. Imagine malware telling your security software, “Nothing suspicious here. Move along.”

It sounds absurd, but that is essentially what the researchers demonstrated.

A Poem That Controls Infected Computers

A separate report from The Register described malware called PoeLLM, which compromised thousands of servers, including systems running AI-related services.

The particularly interesting detail was how the malware communicated with its operators. Instead of using an obvious command server address, it retrieved a seemingly harmless poem from GitHub. Certain words in the poem concealed information directing infected computers to their control server. Changing the poem could redirect the malware without modifying the malicious program itself.

This wasn’t a prompt-injection attack. However, it demonstrated how ordinary-looking text can conceal information used for malicious purposes.

What Happens When Criminals Combine These Techniques?

Here’s where things get interesting. Businesses are increasingly using AI to analyze documents, review emails, summarize security alerts, and automate everyday operations. Now imagine an attacker introducing a malicious file containing instructions designed to manipulate the AI analyzing it. That same attack could use innocent-looking documents or publicly hosted content to conceal additional information.

An AI assistant might be persuaded to dismiss a legitimate security warning, overlook suspicious activity, or make a recommendation based on deliberately misleading information. The attacker isn’t necessarily trying to break the AI. They’re trying to make it work for them.

The two reported incidents do not establish that these techniques have been combined in a single attack. However, they demonstrate capabilities that could complement each other. And as organizations give AI access to more information and greater authority to act, the potential consequences increase.

What You Can Do

Organizations don’t need to abandon AI, but they should treat it as another system requiring appropriate cybersecurity protections.

• Know where AI is being used. Identify AI assistants, automated processes, and services connected to business data.
• Limit AI permissions. Don’t give an AI system access to information or capabilities it doesn’t need.
• Protect AI infrastructure. Keep locally hosted AI systems and supporting applications updated, properly configured, and monitored.
• Treat external content as untrusted. Information from documents, websites, and emails should not automatically become instructions for AI.
• Maintain human oversight. Important security, financial, and operational decisions should not depend entirely on AI conclusions.
• Keep traditional defenses. AI complements antivirus, endpoint protection, firewalls, and security monitoring. It doesn’t replace them.

How CDML Can Help

At CDML Computer Services, we help organizations adopt new technologies while maintaining appropriate cybersecurity protections. Whether you’re using cloud-based AI services, automating business processes, or exploring locally hosted AI models, security needs to be part of the implementation.

CDML can help evaluate your technology environment, manage access permissions, protect endpoints and networks, and implement the monitoring and security controls needed to reduce unnecessary exposure.

Final Thoughts

Cybercriminals have always tried to deceive people and bypass security systems. Now they’re learning to manipulate the artificial intelligence that organizations increasingly rely upon. AI can be a powerful tool, but it shouldn’t be trusted blindly, especially when processing information that an attacker might control.

If your organization is expanding its use of AI, contact CDML Computer Services to make sure your cybersecurity strategy is keeping pace.

Stay safe. Stay informed. Stay compliant.

📞 Contact us here: https://cdml.com/contact/
📚 Read more on our blog: https://cdml.com/blog-2
📺 Listen to our blogcasts: https://www.youtube.com/@CDMLComputerServices