How the attack works
An attacker creates a skill package for an AI agent platform and embeds malicious code inside it, often obfuscated with base64 encoding or hidden in a password-protected ZIP to dodge static scanners. The package is published to a skill marketplace or repository, sometimes as one of many skills mass-produced by the same actor. A user or an agent installs the skill, believing it adds legitimate functionality. On execution, the hidden code runs — documented cases include delivering the AMOS infostealer and exfiltrating authentication tokens to achieve remote code execution.
Why it matters
An organization that installs a compromised skill can have credentials and auth tokens stolen, malware run on its systems, and remote access handed to the attacker — one confirmed case affected over 40,000 instances of a single agent platform.
What you can do
- Review the source and publisher reputation of any AI agent skill before installing it, and prefer skills from known, verified publishers.
- Inspect skill contents (SKILL.md and bundled scripts) for obfuscated code, base64 blobs, or password-protected archives before use.
- Restrict which users or systems can install new skills, and require a review step for third-party or community-published skills.
- Monitor installed skills for unexpected network calls, credential access, or execution of encoded/decoded payloads after installation.
Known benign look-alikes
- Legitimate installer scripts from known publishers (brew, nvm, rustup)
- Skills that discuss credential management as documentation
- Security training materials that describe archive extraction with passwords in an educational context (e.g., CTF walkthroughs mentioning 'password = infected123' near the word 'extract')