Security researchers are highlighting the risks associated with AI agent skills, which are often distributed via public repositories. A report by Snyk in February 2026 analyzed 3,984 skills and found that 36.8% contained security issues, with 13.4% classified as critical. Researchers confirmed 76 malicious payloads, some of which remained active at the time of publication.
The vulnerability stems from the fact that agent instructions, often written in plain text, can execute commands with the user's system privileges. Attackers use techniques like base64-encoded commands, environment variable exfiltration, and remote script execution to compromise systems. A notable real-world case, CVE-2025-6514 in the mcp-remote package, affected over 437,000 installations.
To address these threats, the project worklore.dev introduced 'skill-xray,' an open-source tool licensed under MIT. Instead of providing a binary 'safe' or 'unsafe' verdict, the tool categorizes skills into five tiers based on their capabilities, such as network access, file system interaction, or persistence. It generates a report linked to a specific content hash to prevent bait-and-switch attacks.
The tool operates in two layers: a mechanical scanner that performs static analysis without executing code, and an agent-based layer that provides human-readable context. This approach aims to help users assess the 'blast radius' of a skill before execution. Other security solutions, including those from Cisco and Kaspersky, also focus on detecting AI-related malware, with Kaspersky reporting over 15,000 malicious samples identified in a year.
