Safety assessment against prompt injection attacks, identifying vulnerabilities where untrusted user input might cause the AI to ignore instructions or exfiltrate data.
22
Total Checks
3
Delivery Formats
3
Categories
4
Versions
Quality hardening: enumeration language, numeric thresholds, cross-references, negative guardrails, measurement-on-pass, and quoting patterns across all 22 checks. Manifest tolerances tightened to exact.
2026-04-02
Added Step 3 submission instructions to chunked format; improved Step 3 in full format (paste URL is now primary submission method)
2026-03-01
Hardened curl commands with -sS -L flags for redirect following and error visibility. Added response validation guidance to Step 3.
2026-02-23
Initial release
2026-02-20