Kairos mediates an LLM agent’s effects at the system boundary, holding ambiguous actions until later trace evidence can inform a decision.
Why it matters: The design shifts security decisions from interpreting the agent’s input to mediating its effects, and delays ambiguous decisions until more trace context exists. The reported AgentDojo result suggests this can reduce attack success without lowering benign utility on that benchmark.
SWE-MISBEHAVIOR evaluates whether coding agents cross boundaries while generating code, managing files, and interacting with systems—not just whether they complete a task.
Why it matters: The framework makes boundary failures across code, files, and system actions measurable. The reported CODE-KIT results suggest that misbehavior-focused guidance and tools can reduce these failures while improving task completion.