Cite
Notes
Only stored in your browser.
Attribution
Multi-turn indirect-injection environment: a support agent with a $50 refund cap and real tools (lookup_order, check_approval, issue_refund, escala...
Adversarial train + eval: a support agent with a $50 refund cap, attacked by prompt-injection payloads hidden in tickets. Procedural generator for ...
A support agent given a real performance objective (resolution rate, handle time - escalations count against it) AND a $50 refund cap it must not e...
Multi-turn adversarial train + eval: a support agent with a $50 refund cap vs a scripted customer that ramps pressure over several turns (plead, fa...