Security researchers have exposed critical weaknesses in the scanners used to detect malicious add-ons for AI coding agents, demonstrating that simple obfuscation techniques can bypass existing defenses. A team from Hong Kong University of Science and Technology found that their most effective evasion method successfully fooled every scanner tested in more than 90% of attempts, raising concerns about the security of AI development tools that are rapidly gaining adoption in software engineering workflows.
AI coding agents have become increasingly popular tools that assist developers by suggesting code, automating tasks, and integrating with development environments through extensible skill systems. These add-on skills expand agent capabilities but also create attack surfaces where malicious actors can inject harmful code. Current security measures rely on static analysis scanners to identify threats before skills are deployed, but the research reveals these defenses are inadequate against even basic evasion techniques.
The researchers developed SkillCloak, a framework that applies straightforward modifications to malicious skills while preserving their harmful functionality. These changes are simple enough that they do not require sophisticated technical knowledge, yet they consistently defeat commercial and open-source scanning solutions. The techniques work by altering code structure and presentation in ways that confuse pattern-matching algorithms without affecting the underlying malicious behavior when the code executes.
The implications extend beyond individual developers to organizations deploying AI coding assistants across engineering teams. Compromised skills could exfiltrate proprietary code, inject backdoors into production systems, or manipulate development pipelines. As companies integrate these AI tools more deeply into their workflows, the attack surface expands, and the potential impact of successful exploits grows proportionally.
In response to the vulnerabilities they uncovered, the research team developed a runtime checker that monitors skill behavior during execution rather than relying solely on static analysis. This dynamic approach successfully detected most evasion attempts in testing. Organizations using AI coding agents should implement layered security controls, including runtime monitoring, strict permission models for skills, and regular audits of installed add-ons until scanner technology improves to match evolving evasion techniques.
Source: https://thehackernews.com/2026/07/new-skillcloak-technique-lets-malicious.html