Skip to content

Safety Net Analyzer

Analyzes defense-in-depth coverage by mapping safety controls to threat categories. Identifies redundancies, single points of failure, and blind spots.

CLI

# Full analysis with all controls
python -m replication safety-net

# Analyze specific controls only
python -m replication safety-net --controls drift,killchain,escalation

# Set minimum desired depth
python -m replication safety-net --min-depth 3

# JSON output
python -m replication safety-net --format json

# HTML report
python -m replication safety-net --format html -o safety_net.html

# List available controls
python -m replication safety-net --list-controls

# List threat categories
python -m replication safety-net --list-threats

Programmatic

from replication.safety_net import SafetyNetAnalyzer, analyze_safety_net

# Quick analysis
report = analyze_safety_net()
print(report.render())
print(f"Score: {report.coverage_score}/100")

# Custom analysis
analyzer = SafetyNetAnalyzer(
    controls=["drift", "killchain", "escalation", "canary"],
    min_depth=3,
)
report = analyzer.analyze()

# Check results
for blind_spot in report.blind_spots:
    print(f"UNCOVERED: {blind_spot.category}")

for spof in report.spofs:
    print(f"SPOF: {spof.category} → {spof.controls[0]}")

Coverage Score

The score (0–100) reflects defense-in-depth quality:

Depth Points
0 (blind spot) 0%
1 (SPOF) 40%
2 80%
3+ 100%

Threat Categories

  • reconnaissance — Probing environment, testing boundaries
  • resource_abuse — Hoarding compute, memory, storage
  • privilege_escalation — Gaining broader permissions
  • lateral_movement — Reaching other agents/services
  • defense_evasion — Avoiding detection, stealth
  • command_control — Establishing covert channels
  • data_exfiltration — Unauthorized data extraction
  • self_replication — Creating unauthorized copies
  • goal_drift — Subtle deviation from objectives
  • deception — Misrepresenting state or intentions
  • supply_chain — Compromising dependencies
  • prompt_injection — Manipulating via injected prompts