{
 "license": "CC-BY-4.0",
 "attribution": "Fide AI, Agentic Cyber Explorer",
 "url": "https://agentic-cyber-explorer.pages.dev/events/injecagent-benchmark-2024/",
 "asOf": "2026-09-26",
 "id": "injecagent-benchmark-2024",
 "date": "2024-03-05",
 "datePrecision": "day",
 "title": "InjecAgent benchmarks indirect prompt injection against tool-integrated LLM agents",
 "lane": "defense",
 "kind": "benchmark",
 "summary": "Zhan, Liang, Ying and Kang release InjecAgent, a benchmark of 1,054 test cases spanning 17 user tools and 62 attacker tools, covering direct harm to users and exfiltration of private data. They evaluate 30 LLM agents and find a ReAct-prompted GPT-4 agent vulnerable in about a quarter of cases.",
 "whyItMatters": "It was an early systematic measurement showing that tool-using agents follow instructions embedded in tool outputs.",
 "actors": [
  "uiuc"
 ],
 "topics": [
  "prompt-injection",
  "tool-and-mcp-security",
  "data-exfiltration"
 ],
 "atlas": [
  "untrusted-content",
  "tools"
 ],
 "artifacts": [
  "injecagent",
  "gpt-4-family"
 ],
 "sources": [
  {
   "url": "https://arxiv.org/abs/2403.02691",
   "publisher": "arXiv",
   "title": "InjecAgent: Benchmarking Indirect Prompt Injections in Tool-Integrated Large Language Model Agents",
   "date": "2024-03-05",
   "type": "primary",
   "accessed": "2026-09-25"
  }
 ],
 "keyFacts": [
  {
   "fact": "Benchmark has 1,054 test cases, 17 user tools and 62 attacker tools; 30 LLM agents evaluated.",
   "locator": "Abstract"
  },
  {
   "fact": "ReAct-prompted GPT-4 was vulnerable to the attacks 24% of the time; a reinforced 'hacking prompt' nearly doubled attack success.",
   "locator": "Abstract"
  }
 ],
 "significance": 3,
 "fideQuestions": [],
 "methods": [
  "agent-data-exfiltration",
  "indirect-prompt-injection"
 ],
 "review": "assistant-drafted",
 "addedOn": "2026-09-25"
}