{
 "license": "CC-BY-4.0",
 "attribution": "Fide AI, Agentic Cyber Explorer",
 "url": "https://agentic-cyber-explorer.pages.dev/events/google-gemini-4-carbon-cyber-evaluations-2026/",
 "asOf": "2026-10-30",
 "id": "google-gemini-4-carbon-cyber-evaluations-2026",
 "date": "2026-10-28",
 "datePrecision": "day",
 "title": "Google announces Gemini 4 Carbon cyber results and phased access for defenders",
 "lane": "capability",
 "kind": "eval-report",
 "summary": "Google announces Gemini 4 Carbon with vendor-reported scores of 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, evaluated without guardrails. It describes phased access for selected cyber defenders through Fairwind, with broader product availability planned after further safeguard work.",
 "whyItMatters": "It adds model-specific cyber measurements and access conditions while distinguishing evaluation capability from deployed availability.",
 "actors": [
  "google-deepmind"
 ],
 "topics": [
  "capability-evaluation",
  "vulnerability-discovery",
  "exploit-development",
  "vulnerability-repair",
  "access-controls"
 ],
 "atlas": [
  "model",
  "access-gate",
  "eval-environment"
 ],
 "artifacts": [
  "gemini",
  "sec-bench-pro",
  "exploitgym",
  "cwe-bench"
 ],
 "sources": [
  {
   "url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/",
   "publisher": "Google DeepMind",
   "title": "Gemini 4 Carbon: our next frontier model",
   "date": "2026-10-28",
   "type": "primary",
   "accessed": "2026-10-30"
  },
  {
   "url": "https://arxiv.org/abs/2605.26548v2",
   "publisher": "Hwiwon Lee et al. (arXiv)",
   "title": "SEC-bench Pro: Can Language Models Solve Long-Horizon Software Security Tasks?",
   "date": "2026-07-20",
   "type": "primary",
   "accessed": "2026-10-30"
  }
 ],
 "keyFacts": [
  {
   "fact": "Google labels the three scores as discovery, exploitation and patching. SEC-bench Pro’s authors describe known-vulnerability tasks that give source paths, vulnerability class and expected error while withholding the reference proof of concept and root-cause location. Its score is not a blind unknown-vulnerability discovery rate; Google’s announcement does not establish identical harness or grading settings.",
   "locator": "Google, State-of-the-art cybersecurity performance; SEC-bench Pro v2, Section III-D and Figure 2"
  },
  {
   "fact": "Google reports 0.0% attack success on the evaluated Gray Swan indirect prompt-injection benchmark. That benchmark result does not establish resistance to all adaptive attacks.",
   "locator": "Advancing frontier model safety & security"
  },
  {
   "fact": "The announcement places general developer, business and consumer access in the future. It does not establish that every applicant has access or that all product settings match unguarded evaluations.",
   "locator": "Coming soon to developers, businesses, and consumers"
  }
 ],
 "significance": 4,
 "fideQuestions": [
  "FID-075",
  "FID-088"
 ],
 "methods": [
  "ai-assisted-exploitation",
  "automated-patching",
  "capability-restriction",
  "ctf-benchmarks",
  "indirect-prompt-injection"
 ],
 "review": "assistant-drafted",
 "addedOn": "2026-10-30",
 "revisions": []
}