Speaker
Description
LLM agents that browse, read email, edit files and run commands act on whatever text they read. This exposes them to indirect prompt injections: instructions hidden in a document, a web page or an email that hijacks the agent into doing something the user never asked for. We build a sandboxed agent with eleven simulated tools (file system, inbox and outbox, calculator, an allow-listed terminal and a local browser), served by open models through Ollama, and a suite of 51 injection scenarios that cover three attacker goals (data exfiltration, unauthorized action, destruction), four techniques (direct, role-play, encoded, hidden markup) and three delivery channels (email, file, web page). Three runtime filters sit between the agent and its tools: a keyword blacklist over tool arguments, an allow-list derived automatically from the entities named in the user's task, and an LLM judge that sees both the task and the proposed call. In the evaluation we separate compromise, where the agent attempts the injected action, from execution, where the action goes through, which credits a defense for stopping an already hijacked agent. On qwen3:14b, the undefended agent attempts the injected action in 10% of scenarios (28 of 204 attempts across all runs) and completes 78% of benign tasks; role-play injections are by far the most effective technique (32% of attempts), while encoded ones never succeed. The keyword filter stops only 2 of 9 attempted actions, the allow-list 6 of 7 and the LLM judge 5 of 7; they block at least one legitimate call on 12%, 12% and 10% of benign tasks, and the judge adds 1.2 s per task (1.1x latency). The attacks that get through are those that abuse the task itself, such as deleting a file the user named or leaking a secret inside an otherwise legitimate email. We release the benchmark, the harness and all traces.