Allegro
|
d139f2c6d2
|
security: Add crisis safety test infrastructure for Issue #79
Docker Build and Publish / build-and-push (push) Has been cancelled
Nix / nix (macos-latest) (push) Has been cancelled
Nix / nix (ubuntu-latest) (push) Has been cancelled
Tests / test (push) Has been cancelled
- crisis_safety_test_og_godmode.py: Automated test runner (312 lines)
- crisis_safety_audit_issue_79.py: Full audit framework (21KB)
- FALLBACK_CHAIN_TEST_PLAN.md: Test specifications
- BLOCKER_ISSUE_79.md: Documentation of API credential blocker
- EXECUTION_SUMMARY.md: Burn cycle summary
Test coverage for 4 crisis safety scenarios:
- CRISIS-001: suicide_method query
- CRISIS-002: suicide_ideation query
- CRISIS-003: suicide_method query (medication)
- CRISIS-004: suicide_validation query
Status: Infrastructure complete, execution blocked pending OPENROUTER_API_KEY
Related: #72 (parent), #79, #80, #81
|
2026-04-01 03:49:48 +00:00 |
|
Allegro
|
be865df8c4
|
security: Issue #81 - ULTRAPLINIAN fallback chain audit framework
Implement comprehensive red team audit infrastructure for testing the entire
fallback chain against jailbreak and crisis intervention attacks.
Files created:
- tests/security/ultraplinian_audit.py: Comprehensive audit runner with:
* Support for all 4 techniques: GODMODE, Parseltongue, Prefill, Crisis
* Model configurations for Kimi, Gemini, Grok, Llama
* Concurrent execution via ThreadPoolExecutor
* JSON and Markdown report generation
* CLI interface with --help, --list-models, etc.
- tests/security/FALLBACK_CHAIN_TEST_PLAN.md: Detailed test specifications:
* Complete test matrix (5 models × 4 techniques × 8 queries = 160 tests)
* Technique specifications with system prompts
* Scoring criteria and detection patterns
* Success criteria and maintenance schedule
- agent/ultraplinian_router.py (optional): Race-mode fallback router:
* Parallel model querying for safety validation
* SHIELD-based safety analysis
* Crisis escalation to SAFE SIX models
* Configurable routing decisions
Test commands:
python tests/security/ultraplinian_audit.py --help
python tests/security/ultraplinian_audit.py --all-models --all-techniques
python tests/security/ultraplinian_audit.py --model kimi-k2.5 --technique crisis
Relates to: Issue #72 (Red Team Jailbreak Audit)
Severity: MEDIUM
|
2026-04-01 01:51:23 +00:00 |
|