Malware Injection Threats on the Rise
Sergey Lozhkin, the head of Kaspersky’s global research and analysis team for the Asia-Pacific, Middle East, Turkey, and Africa, recently highlighted a concerning trend in cybersecurity: the increasing prevalence of malware injection attacks targeting AI agents. During a Kaspersky cybersecurity event held in Tbilisi, Georgia, Lozhkin emphasized that these AI-driven threats are increasingly emerging from outside an organization’s network, with application programming interfaces, frameworks, plugins, and the agents themselves presenting direct risks.
Lozhkin pointed out that traditional signature-based analysis, which relies on identifying direct commands, is becoming obsolete. “Currently, when you or your agent activates a skill, it appears entirely legitimate, performing its intended function. However, since everything operates in the cloud, these skills possess commands that connect to external resources and retrieve instructions,” he explained.
Emerging Cybersecurity Risks from Autonomous AI Agents
Despite these challenges, organizations are not without defenses. Dmitry Galov, who leads Kaspersky’s global research and analysis team for Russia and the CIS, shared insights on effective strategies for combating AI malware. He noted that established tools like behavioral analysis are crucial for detecting malicious code at the moment it executes.
Another viable strategy involves sandboxing AI applications, allowing them to operate freely within a controlled environment while restricting their interactions with the broader internet. “Organizations need not prohibit employees from utilizing AI due to potential dangers; instead, they can implement guardrails to ensure sensitive data remains protected,” Galov advised.
Nonetheless, the emergence of new attack vectors signifies a fundamental shift in how organizations must approach security. With malicious code capable of lying dormant behind seemingly legitimate skills until it reaches the cloud, Lozhkin argued that merely scanning a file once and trusting the outcome is insufficient. The rapid pace at which skills and plugins are published further complicates the situation, outpacing the old security model.
“It’s imperative to think critically and implement robust defensive mechanisms, as the volume of skills is growing exponentially. We are entering an era where security researchers, antivirus vendors, and companies must collaboratively devise innovative solutions,” he cautioned. “Otherwise, I have some bad news for you.”
- The writer traveled to Tbilisi, Georgia as a guest of Kaspersky.
