🔬 New research: when an LLM is asked to "save this text verbatim," inference-framework parsers reinterpret embedded protocol markers as live tool-call control structure — silently injecting, rewriting, or dropping tool calls.
In our end-to-end lab, an agent with a shell tool received and executed arbitrary commands — triggered by a single sentence. The model recognized the injection, copied the payload anyway, and the server-side parser delivered the call. The agent never issued it.
We also showed tool_choice enforcement can be bypassed: server forces "only call tool X," parser delivers tool Y.
Full writeup: https://t.co/qqd7ZTeNA8
Research by SEC4SI Team
@ProcessorLee · @VillanCh · @bestswngs · @chun_springX · @bbznop · @0chencc