What each coding agent can enforce
Seven coding agents, 16 questions each. Every agent can block a tool call with a hook, but sandboxes are mostly off by default and a “deny” rule often only means “ask”.
Research
What we find when we point our open-source tools at public code. Each report comes with its method and data.
Seven coding agents, 16 questions each. Every agent can block a tool call with a hook, but sandboxes are mostly off by default and a “deny” rule often only means “ask”.
38% of them are exact copies, and none hid an instruction. The real risk is skills that switch the agent’s safety off on purpose.