Liran Tal flags agent skill security risks

Blindly installing third-party AI agent skills is the modern equivalent of executing arbitrary curl-to-bash scripts, opening up a massive, under-secured supply-chain playground for attackers.

• –Untrusted Execution: Agent skills define instructions and tool capabilities that run in developer environments, creating potential routes for malicious execution.
• –Blind Trust: Developers frequently import skills without inspecting their underlying prompt instructions, configuration metadata, or capabilities, creating highly vulnerable blindspots.
• –Expanded Attack Surface: As AI coding assistants are granted deep filesystem and tool access, compromised skills can lead to seamless data exfiltration, credential theft, and persistent malware installation.