Findings/rl-defenders-beat-llm-defenders

In the CAGE 4 simulation, reinforcement-learning defenders outperformed LLM-based defenders.

Revalidatemeasured1 evidence record from 1 independent sourceassistant-drafted

Scope: what this does not show

One simulation designed for RL agents, with 2 episodes per scenario. The all-LLM team used GPT-4o-mini only; other early-2025 models were tested only as one member of mixed teams.

Revalidate: Older than its half-life with no newer evidence. May no longer hold.

Evidence

May 7, 2025

UC Santa Cruz study integrates LLM agents into CAGE 4 and finds RL defenders still outperform them

arXiv

Status history

2025-05-07ReportedUC Santa Cruz study. · record
2026-09-26RevalidateComputed: 507 days since the last evidence, past the 365-day half-life.

What would change this: A comparison with current models.
Evidence kind: measured: An experiment or evaluation produced the number.
Freshness: Evidence 507d old · past its 365-day recheckNewest evidence May 7, 2025. Half-life 365 days.
Methods tested: Cyber ranges and simulations
Topics: Autonomous defense
Fide research: FID-076: Authorization boundaries and AI control in cybersecurity
FID-087: Collective resilience in autonomous cyber defense

Cite this record

Fide AI. "In the CAGE 4 simulation, reinforcement-learning defenders outperformed LLM-based defenders.." Agentic Cyber Explorer, record rl-defenders-beat-llm-defenders. https://agentic-cyber-explorer.pages.dev/findings/rl-defenders-beat-llm-defenders/ (corpus as of 2026-09-26).

BibTeX

@misc{fide-ace-rl-defenders-beat-llm-defenders,
  author = {{Fide AI}},
  title = {In the CAGE 4 simulation, reinforcement-learning defenders outperformed LLM-based defenders.},
  howpublished = {Agentic Cyber Explorer},
  year = {2026},
  url = {https://agentic-cyber-explorer.pages.dev/findings/rl-defenders-beat-llm-defenders/},
  note = {Record rl-defenders-beat-llm-defenders; corpus as of 2026-09-26}
}

Data: JSON · CC BY 4.0

Know of evidence that changes this? Tell us.