Skip to content

English · Español

Phase 37 — Quizzes

🇪🇸 Espejo legible del banco de preguntas; la fuente canónica es data/quizzes/phase-37-security-safety.yaml.

q-37-01 — Defense layer that scores "high" against every injection category

  • Boundary marking on user input
  • System prompt placed at the end of context
  • Ground-truth verification against the §A13 conjugation table ← correct
  • A behavioral output classifier

q-37-02 — Pick every prompt-injection category named in the §A13 taxonomy

  • Direct override ← correct
  • Role substitution ← correct
  • Authority claim ← correct
  • Hypothetical framing ← correct
  • Indirect / data-channel ← correct

q-37-03 — Why is ground-truth verification feasible at §A13 scale?

Free response. Acceptable answers contain scope.

The §A13 output space is closed and small (~600 forms); a finite table can enumerate every correct answer.

q-37-04 — Canonical project threat model file?

  • docs/phase-37-security-safety/README.md
  • security/THREATS.md ← correct
  • LYNX_CORTEX.md §10
  • CLAUDE.md §0

See theory/06-prompt-injection-taxonomy-a13-examples.md and the canonical security/THREATS.md.