PROMPT INJECTION THROUGH TOOL OUTPUT

Summary from the TLDR AI newsletter

Prompt injections hidden in tool results can evade conventional safeguards because input and action screens inspect separate moments of an agent loop. The proposed signal is a “precedent gap,” where an agent suddenly makes a tool call or uses arguments absent from its execution history.