{
 "license": "CC-BY-4.0",
 "attribution": "Fide AI, Agentic Cyber Explorer",
 "url": "https://agentic-cyber-explorer.pages.dev/events/deepmind-adaptive-injection-agent-swarms-2026/",
 "asOf": "2026-11-17",
 "id": "deepmind-adaptive-injection-agent-swarms-2026",
 "date": "2026-11-09",
 "datePrecision": "day",
 "title": "DeepMind tests adaptive injection swarms against synthetic tool-call defenses",
 "lane": "capability",
 "kind": "paper",
 "summary": "A Google DeepMind preprint tests collaborative attacker agents on a synthetic instruction-hierarchy benchmark with randomly selected, non-malicious tool-call goals. Stronger attackers and larger query budgets expose failures that static tests miss. The experiments isolate model behavior and do not measure successful real-world data theft or a complete deployed defense stack.",
 "whyItMatters": "It shows why robustness claims depend on adaptive attacker resources and on which safeguards are actually included in an evaluation.",
 "actors": [
  "google-deepmind"
 ],
 "topics": [
  "prompt-injection",
  "capability-evaluation",
  "eval-validity",
  "multi-agent-security"
 ],
 "atlas": [],
 "artifacts": [
  "gemini"
 ],
 "sources": [
  {
   "url": "https://arxiv.org/abs/2611.11861v1",
   "publisher": "Google DeepMind / arXiv",
   "title": "Scaling Adaptive Prompt Injection Attacks with Agent Swarms",
   "date": "2026-11-09",
   "type": "primary",
   "accessed": "2026-11-15"
  }
 ],
 "keyFacts": [
  {
   "fact": "The dataset asks whether untrusted tool output can induce a random tool call unrelated to the legitimate user request. Attack goals are deliberately non-malicious to isolate instruction-hierarchy behavior.",
   "locator": "Sections 1–3"
  },
  {
   "fact": "The Agentic Optimizer uses collaborating agents to search adaptively; results vary with attacker capability and the defender-query budget.",
   "locator": "Agentic Optimizer; experimental results"
  },
  {
   "fact": "The Gemini experiments disable safety classifiers for both attacker and defender models, as well as outside-model prompt-injection mitigations. Results do not evaluate all safeguards of a deployed Gemini product.",
   "locator": "Section 5.1: evaluation setup"
  },
  {
   "fact": "The simplified next-tool-call setting does not measure a full intrusion, exfiltration or multi-step production compromise.",
   "locator": "Dataset design and limitations"
  }
 ],
 "significance": 4,
 "fideQuestions": [
  "FID-075",
  "FID-087"
 ],
 "methods": [
  "adaptive-red-teaming",
  "compute-scaled-evaluation",
  "ctf-benchmarks",
  "indirect-prompt-injection"
 ],
 "review": "assistant-drafted",
 "addedOn": "2026-11-15",
 "revisions": []
}