feat: Add AgentHealthBench — agent behavioral health probes and detectors

[Gach0ng]

Summary

Adds AgentHealthBench, a 4-tier agent behavioral health testing suite
integrated as garak probes and detectors.

4-Tier Risk Model

Tier	Probe	Prompts	Detectors
Atomic	agent_health_atomic.ToolMisuse	132	agent_health_action_trace.ActionTraceDetector
Atomic	agent_health_atomic.ActionChain	109	(same)
Comm	agent_health_comm.IdentitySpoof	113	(same)
Emergent	agent_health_emergent.CascadeFailure	120	(same)
Immunity	agent_health_immunity.VaccineRecall	123	(same)

Total: 597 prompts across 5 probes.

New Detectors (4)

• agent_health_action_trace.ActionTraceDetector — 8 blacklist regexes
• agent_health_permission_check.PermissionCheckDetector — 14 permission regexes
• agent_health_identity_verify.IdentityVerifyDetector — 12 identity-compliance regexes
• agent_health_cascade_detect.CascadeDetector — 14 cascade-failure regexes

New Post-Processor

• agent_health_health_scorer.HealthScorer — CLI that reads garak .report.jsonl
  and outputs a weighted 0–100 health score with per-tier breakdown.

Tests

• 35 probe tests (load, attrs, prompt count ≥ 30, lang=en, active, primary_detector)
• 22 detector tests (load, hit/no-hit for each detector)
• 57/57 passing

Usage

garak --model_type openai --model_name gpt-4 \
      --probes agent_health_atomic.ToolMisuse,agent_health_atomic.ActionChain,agent_health_comm.IdentitySpoof,agent_health_emergent.CascadeFailure,agent_health_immunity.VaccineRecall \
      --detectors agent_health_action_trace.ActionTraceDetector

python garak/detectors/agent_health_health_scorer.py path/to/report.jsonl

[jmartin-tech]

Garak is not a benchmark and while some prompts from existing benchmarks are accepted probes that do not implement a specific adversarial attack technique need strong reasons to be included.

Some early feedback not a full review of viability for acceptance of these probes.

Probe and detector names with common grouping are currently organized into single module files with multiple classes included to group them. #1503 represents an desire to refactor for the overall project to allow for various plugins to be individual files, however that effort is in queue and not in scope for this PR.

Please match the code organization to align probe/detector classes something like:

detectors.agent_health.ActionTrace
detectors.agent_health.Cascade
detectors.agent_health.IdentityVerify
detectors.agent_health.PermissionCheck

probes.agent_health.*

Also many of the prompts included here have coverage in various existing probes. Spending extra inference during testing to confirm these duplicate prompts has limited value.

[jmartin-tech]

This is not a detector it is either an independent tool or an analysis tool and should either be in the tools path for external use or in the garak/analyze module path for distribution usage that the pip installed package.