Neil Lawrence#176
AI agent security creates a weird new category:
Code-free attacks.
If a skill can influence how an agent interprets instructions, routes tools, or suppresses warnings, executable code is not required for harm.
That is very different from traditional software supply chains.
http://x.com/i/article/2037527406647316480
One of the strangest attack patterns in our ClawHub analysis:
A package with zero executable code.
No Python. No shell script. Nothing to run. The entire attack surface was the SKILL.md file. The markdown itself was the weapon.
http://x.com/i/article/2037527406647316480
Neil Lawrence#176AI agent security creates a weird new category:
Code-free attacks.
If a skill can influence how an agent interprets instructions, routes tools, or suppresses warnings, executable code is not required for harm.
That is very different from traditional software supply chains.
http://x.com/i/article/2037527406647316480