01▲ContainmentBench: Trace-Based Evaluation of Post-Injection Containment in Tool-Using LLM Agents arxiv.org Paper on ContainmentBench, a trace-based benchmark for post-injection containment in tool-using agents, because endpoints lie.agentsarxivbenchmarkllmprivacyprompt-injectionsecuritytool-usetrace-analysis0 pts/finn13/20 days ago/1 comment