Voicebot red-team suite
Forty-five scripted conversations and an adaptive attacker try to break a voice receptionist
Each script is a short caller conversation, from prompt-extraction attempts to fake-authority tricks, replayed in text against the receptionist's prompt and tools and scored by fixed checks. No model grades the results. Watch the final full run play back case by case.
Adaptive attacker
A second model plays the caller and changes tactics as it goes
Ten goals, two attempts each, up to six turns per attempt. An attempt ends as held, as broken, or with the attacker declining to play.
Checking the checks
The harness checks its own checks
The self-test feeds 36 fabricated transcripts through the checks. Each bad one must be flagged and each good one must stay clean. This is its real output.
Source: the red-team result files and the harness self-test output from 22 October 2026, replayed here as metadata only: case ids, categories, sample counts, pass or fail, check names, goal names, outcomes, turn counts and call totals.