{
 "license": "CC-BY-4.0",
 "attribution": "Fide AI, Agentic Cyber Explorer",
 "url": "https://agentic-cyber-explorer.pages.dev/events/ucsc-llms-autonomous-cyber-defenders-cage4-2025/",
 "asOf": "2026-09-26",
 "id": "ucsc-llms-autonomous-cyber-defenders-cage4-2025",
 "date": "2025-05-07",
 "datePrecision": "day",
 "title": "UC Santa Cruz study integrates LLM agents into CAGE 4 and finds RL defenders still outperform them",
 "lane": "defense",
 "kind": "paper",
 "summary": "Researchers led by UC Santa Cruz integrated LLM agents into the CybORG CAGE 4 multi-agent defence environment and proposed a communication protocol for mixed LLM and RL teams. In their runs an all-RL team scored far better reward than an all-LLM (GPT-4o-mini) team and acted about 104 times faster, though the authors highlight LLM explainability and note the environment was designed for RL agents.",
 "whyItMatters": "The authors describe it as the first study of LLM agents in a multi-agent autonomous cyber defense environment, and it cautions against assuming LLMs beat trained RL policies.",
 "actors": [
  "uc-santa-cruz"
 ],
 "topics": [
  "autonomous-defense",
  "multi-agent-security"
 ],
 "atlas": [
  "eval-environment",
  "other-agents"
 ],
 "artifacts": [
  "cage-challenge-4",
  "deepseek",
  "gpt-4-family",
  "openai-o-series"
 ],
 "sources": [
  {
   "url": "https://arxiv.org/abs/2505.04843",
   "publisher": "arXiv",
   "title": "Large Language Models are Autonomous Cyber Defenders",
   "date": "2025-05-07",
   "type": "primary",
   "accessed": "2026-09-25"
  },
  {
   "url": "https://arxiv.org/html/2505.04843v2",
   "publisher": "arXiv",
   "title": "Large Language Models are Autonomous Cyber Defenders (v2 HTML)",
   "date": "2025-07-19",
   "type": "primary",
   "accessed": "2026-09-25"
  },
  {
   "url": "https://github.com/r4wd3r/llms-are-acd",
   "publisher": "GitHub (r4wd3r)",
   "title": "llms-are-acd: LLM+RL Framework for ACD Agents",
   "type": "primary",
   "accessed": "2026-09-25"
  }
 ],
 "keyFacts": [
  {
   "fact": "Mean reward: all-RL team (KEEP) -493 (sd 95.9) vs all-LLM team with GPT-4o-mini -2547.2 (sd 498.8).",
   "locator": "Section IV-A, Figure 5"
  },
  {
   "fact": "RL agents were about 104.1 times faster at action selection; all-RL runs averaged 45.2 s vs 4704.6 s for all-LLM (GPT-4o-mini).",
   "locator": "Section IV-A"
  },
  {
   "fact": "Models tested: GPT-4o-mini, o3-mini, o1-mini and DeepSeek-V3; 2 episodes of 500 steps per scenario.",
   "locator": "Experimental setup"
  },
  {
   "fact": "Presented at the 2025 IEEE CAI Workshop on Adaptive Cyber Defense.",
   "locator": "arXiv comments"
  }
 ],
 "significance": 3,
 "fideQuestions": [
  "FID-076",
  "FID-087"
 ],
 "methods": [],
 "review": "assistant-drafted",
 "addedOn": "2026-09-25"
}