Skip to content

Latest commit

 

History

History
74 lines (53 loc) · 3.74 KB

File metadata and controls

74 lines (53 loc) · 3.74 KB

Simulation Overestimates Multi-Agent Cascade by 37pp — But Topology Matters More Than We Thought

Zero-trust cuts cascade by ~7pp (not 40pp). Topology DOES matter — hierarchical is protective (0.560 vs flat 0.733). The simulation overestimates severity by 37pp but correctly predicts zero-trust is best.

Blog post: Our Simulation Was Wrong by 37 Percentage Points

govML Quality License

Key Result

Key Results

Agent Count Cascade Rate (mean +/- std) Poison Rate (mean +/- std)
2 1.000 +/- 0.000 0.945 +/- 0.006
3 1.000 +/- 0.000 0.960 +/- 0.011
5 1.000 +/- 0.000 0.974 +/- 0.009
7 1.000 +/- 0.000 0.981 +/- 0.008
10 1.000 +/- 0.000 0.978 +/- 0.006

Core insight: Zero-trust cuts cascade by ~7pp (not 40pp). Topology DOES matter — hierarchical is protective (0.560 vs flat 0.733). The simulation overestimates severity by 37pp but correctly predicts zero-trust is best.

Quick Start

git clone https://github.com/rexcoleman/multi-agent-security
cd multi-agent-security
pip install -e .
bash reproduce.sh

Project Structure

FINDINGS.md # Research findings with pre-registered hypotheses and full results
EXPERIMENTAL_DESIGN.md # Pre-registered experimental design and methodology
HYPOTHESIS_REGISTRY.md # Hypothesis predictions, results, and verdicts
reproduce.sh # One-command reproduction of all experiments
governance.yaml # govML governance configuration
CITATION.cff # Citation metadata
LICENSE # MIT License
pyproject.toml # Python project configuration
scripts/ # Experiment and analysis scripts
src/ # Source code
tests/ # Test suite
outputs/ # Experiment outputs and results
data/ # Data files and datasets
config/ # Configuration files
docs/ # Documentation and decision records

Methodology

See FINDINGS.md and EXPERIMENTAL_DESIGN.md for detailed methodology, pre-registered hypotheses, and full experimental results with multi-seed validation.

Limitations

  • Simulation-based, not real LLM agents — and showed the gap is 48pp. real agent experiments found 49% poison rate where this simulation predicted 97%. The simulation overestimates cascade severity because real agents have inherent semantic resistance. Qualitative findings (zero-trust > implicit) hold but quantitative predictions do not transfer. See FINDINGS for the simulation-to-real gap analysis.
  • Fixed cascade probability. The base cascade probability (0.15) was tuned for differentiation. Real-world cascade probability depends on LLM capability, prompt design, and task complexity. We report relative comparisons between conditions, not absolute rates.
  • No real-time threat intelligence. The simulation doesn't model evolving threats, model updates, or adversary learning over multiple encounters. Each run is a static snapshot.
  • 5 agents maximum in most experiments. E1 goes to 10 agents, but the primary results use 5. Larger systems (50-100 agents) may exhibit different cascade dynamics (e.g., partition effects, natural firebreaks).
  • Single compromised agent assumption. All experiments start with exactly 1 compromised agent. Multi-point compromise (2+ initial attackers) may produce qualitatively different dynamics.

Citation

If you use this work, please cite using the metadata in CITATION.cff.

License

MIT 2026 Rex Coleman


Governed by govML v3.3