Measured, Not Asserted

    Platform Trust Scorecard

    The same adversarial red-team corpus that gates every release runs on a fixed cadence against our deployed guardrails. This is that scorecard, not a marketing claim.

    Overall Detection Rate

    54%

    51/94 adversarial probes detected

    Benign False-Positive Rate

    0%

    0/13 legitimate requests over-blocked

    Last Verified

    less than an hour ago

    Corpus probes-94 · re-checked hourly

    Detection Rate by Category

    Share of adversarial probes in each category that the deployed guardrail stack detected, this run.
    JailbreakPrompt InjectionSystem Prompt LeakHarmful ContentPII ExtractionBias AmplificationHallucination InductionData Extraction0%25%50%75%100%

    Methodology

    Each run scores our deployed input filter (content-safety + prompt-injection detection) against a fixed 94-probe adversarial corpus spanning 8 categories: jailbreak, prompt injection, system-prompt leak, harmful content, PII extraction, bias amplification, hallucination induction, and data extraction, plus a benign-control set to measure over-blocking. This tests whether the filter recognizes an attack pattern in the prompt text itself, before it would ever reach a model. No LLM is called and no model's output is evaluated here. The same corpus and filter gate every release before it ships.

    Pattern-matching reliably catches jailbreak, prompt-injection, system-prompt-leak, and harmful-content attempts. It structurally cannot catch PII extraction, bias, hallucination, or data extraction, since those require judging meaning, not matching text, and will stay near zero until we ship a semantic or model-based rail. We don't publish the individual probes themselves, since that would hand attackers a playbook, but the category names, corpus size, and every measured result are exactly what our own release gate sees. Gaps stay visible on purpose rather than being hidden behind a rounded-up composite number.