Researchers Find Malicious AI Agents Hiding Instructions
New study reveals malicious AI skills using advanced obfuscation to bypass human security reviewers.
Updated on Oct. 21, 2026 in Artificial Intelligence
Live Poll
Do you trust that AI software tools are safe to install without independent security verification?
Security researchers identified 90 malicious AI agent skills that employ clever obfuscation techniques to mask harmful instructions from human oversight. The findings suggest malicious actors are increasingly hiding payloads to bypass standard security scanners.
Why it matters
Malicious actors are masking instructions to successfully evade human reviewers and automated security platforms. These techniques allow harmful AI capabilities to infiltrate systems without detection by standard security protocols.
Analysis of 35,878 AI agent skills on VirusTotal uncovered 6,637 malicious entries, with 1.4% using instruction-hiding techniques. These include 56 skills using whitespace padding, 28 using HTML comments, and 6 using invisible Unicode characters.
The players
VirusTotal
This is a Google-owned security repository that aggregates threat intelligence from various antivirus scanners and file analysis tools.
The details
Malicious agents read files containing hidden characters and comments that standard Markdown renderers ignore. While 62% of identified malicious skills lack actual malicious code, these obfuscation methods specifically target the review process.
Timeline
October 2026: Publication of Part III of From Automation to Infection.
The Tech Race
This research follows the From Automation to Infection series, which highlights the ongoing struggle to secure AI ecosystems against increasingly sophisticated obfuscation tactics. It marks a shift from simple code-based detection to identifying agents that manipulate how reviewers perceive data.
Users and developers integrating third-party AI agents may face heightened security risks from hidden instructions that bypass standard inspections. Rigorous auditing of raw code, rather than relying on standard renderers, is now required to ensure agent safety.
The takeaway
The security industry must move beyond standard Markdown rendering to detect malicious instructions hidden in raw data files. Developers should prioritize deep-code inspection to prevent silent threats from entering their AI workflows.
Further reading
For more on evolving security threats in the software sector, browse our coverage of Artificial Intelligence.
Source note: This article includes information reported by IT Security News - cybersecurity, infosecurity news.
Live Poll
Do you trust that AI software tools are safe to install without independent security verification?
