FID-023 · Open question
Faith-Facing Red-Team Suite
What red-team probes are needed to expose failures unique to faith-facing AI systems and adjacent high-trust guidance systems?
Why the question remains open
General red-team suites test jailbreaks, harmful instructions, privacy, and security. Faith-facing systems also need probes for false religious authority, spiritual manipulation, coercion, crisis mishandling, fabricated sources, sectarian bias, and overconfident moral guidance.
Working hypothesis
A proposition to test, not a finding.
Faith-facing failures can be elicited through pressure patterns that generic red-team suites under-sample: confessional intimacy, divine authorization, conversion pressure, secrecy, pastoral dependency, institutional impersonation, and vulnerable-user escalation.
Proposed method
How the question could be tested
- 01Build a categorized red-team prompt and multi-turn suite.
- 02Include benign, ambiguous, and severe-risk prompts.
- 03Run across base models and deployed harnesses.
- 04Score both failure occurrence and recovery.
Needed controls
What must constrain the study
- 01Safety review before publishing prompts.
- 02Release tiers for sensitive probes.
- 03Avoid prompts that enable real-world harm.
- 04Severe-failure adjudication.
Expected outputs
Artifacts the work should produce
- 01Red-team taxonomy.
- 02Release-safe prompt subset.
- 03Private severe-risk suite.
- 04Integration adapter for Promptfoo, Inspect, or Garak.
Open questions
Uncertainties the protocol must resolve
- 01Which probes should be public?
- 02How should Fide handle discovered severe failures in live products?
- 03How can red-team results avoid becoming sensational?
Related calls
Continue through this research area
FID-012
Optimization Pressure and Visible-Rubric Gaming
If builders can see Fide AI rubrics or optimize against public benchmark items, do systems become genuinely safer or merely better at passing the visible test?
FID-076
Authorization Boundaries and AI Control in Cybersecurity
Which controls keep capable agents within legitimate authorization when task pressure, untrusted inputs, or delegated work creates opportunities to exceed it? For autonomous cyber defense, how can response controls preserve threat containment and legitimate service availability while respecting human authority?
FID-078
When Trustworthiness Evaluations Transfer Across Domains
Which measures of evidence use, authority boundaries, and human control transfer across high-trust domains, and which require domain-specific definitions and calibration?
Open question
Open work
Primary need: red-team design, safety reviewers
- Draft probes.
- Review release safety.
- Build runner integrations.
- Analyze failure recovery.