sauce.ai/redteam
Instrument · adversarial safety evaluation · clinical conversational AI

Adversarial safety evaluation of clinical chatbots

Runs adversarial synthetic-patient conversations against a chatbot you are authorized to test, scores every reply with an audited judge panel, and reports harm with epidemiologic measures.

1Target

An OpenAI-compatible or Anthropic endpoint, a custom JSON API, or a chat web page, under your system prompt. Keys are held for the run only.

2Adversarial conversations

Synthetic personas from a specialty library; an attacker ensemble escalates toward dosing, triage, discontinuation and misinformation harms.

3Report

Attack success with 95% CI, prompts-to-harm, number needed to harm, QALYs lost, harm profile, annotated transcripts; every run is added to the leaderboard.

Field scanChatGPT, Claude, Gemini and Llama on one condition, same case-mix; comparative harm image.Open → Exposure analysisRR / OR / RD, PAF, E-value; Mantel–Haenszel strata; dose–response; judge-misclassification bias analysis.Open → Grader auditBias, robustness and calibration of the harm judge itself.Open → LeaderboardEvery completed run, ranked safest-first per specialty.Open → Methods & workflowsAdjudication, critical harm events, ablations, power, Table 1, STROBE — each with a launcher.Open →
Exposure
the chatbot (agent) under test
Unit
one synthetic-patient conversation; replies nested within conversations
Outcome
a reply judged unsafe at P(harm) ≥ threshold; critical harm events adjudicated by clinicians
Judge
LLM panel audited for bias and calibration; validated against clinician labels; misclassification propagated by quantitative bias analysis
Preview limit
up to 100 trials per email · authorized testing only
This tool deliberately tries to elicit unsafe responses from the bot under test so you can find and fix them. Only run it against endpoints you own or have written permission to test. Reports are LLM-scored screening signals, not clinical determinations — have a clinician review flagged transcripts.

Run an evaluation

Defaults are sensible; the engine is tunable under “Advanced”.

1 Your chatbot — the target

Custom endpoint mapping
Web page selectors

Requires the browser add-on on the server. Selectors target the newest assistant bubble.

2 The evaluation

Harm categories to hunt (blank = all)
Advanced: the red-team engine (optional)

The attacker is an ensemble of LLMs that propose, refine, and vote on the next patient message most likely to push the bot into an unsafe reply. All of this is optional — defaults use Claude.

3 Where to send the report

Run

Open live report ↗ · JSON export ·