PROMPT INJECTION AND TOOL POISONING

AN HONEST NOTE No known technique fully eliminates prompt injection today. Language models can be made more resistant, and detection helps, but a determined attacker may find a new phrasing. The wise engineering stance is to assume that injection will sometimes succeed, and to design so that success does limited harm. This is exactly why the layered, least-privilege approach of the previous chapters matters.