{
 "license": "CC-BY-4.0",
 "attribution": "Fide AI, Agentic Cyber Explorer",
 "url": "https://agentic-cyber-explorer.pages.dev/findings/agent-intrusions-leave-distinctive-traces/",
 "asOf": "2026-11-17",
 "id": "agent-intrusions-leave-distinctive-traces",
 "claim": "Investigators infer agent involvement in intrusions from behavioral traces such as rapid activity bursts, noisy trial and error and natural-language reasoning.",
 "evidenceKind": "observed",
 "scope": "Incident investigations and analyses of exposed attacker infrastructure by security vendors, published from July to October 2026. Sysdig, Wiz and GTIG use these traces to infer that an agent was involved, and TrendAI uses stylistic traces to infer AI assistance; none publishes a validated detection rate, and the traces do not identify the model. TrendAI notes a skilled person could write in a similar style, and GTIG expects stealth-focused actors to be less likely to use publicly available versions of these frameworks.",
 "topics": [
  "ai-enabled-intrusion",
  "threat-intelligence",
  "incident-reporting"
 ],
 "atlas": [],
 "evidence": [
  {
   "event": "sysdig-jadepuffer-agentic-ransomware-2026",
   "note": "Sysdig bases its assessment that an LLM agent drove the intrusion on self-narrating payloads with natural-language reasoning and rapid adaptive retries, including a working fix 31 seconds after a failed login."
  },
  {
   "event": "wiz-agentic-intrusion-fingerprints-2026",
   "note": "Wiz proposes seven behavioral fingerprints from three investigated intrusions, combining sustained variable cadence, reasoning artifacts and adaptive behavior; it gives no validated error rate, and cadence alone does not establish model identity."
  },
  {
   "event": "gtig-mandiant-artex-pentagi-incident-response-2026",
   "note": "From Mandiant case logs, GTIG reports ARTEX execution bursts of roughly 2 to 25 seconds between longer pauses, trial and error far noisier than a competent human operator’s, and recognizable incrementing markers in file names, created account names and email subjects."
  },
  {
   "event": "trendai-three-japan-attack-infrastructures-ai-roles-2026",
   "note": "TrendAI rates one operator’s AI use from stylistic traces in command histories and scripts alone, and says a skilled person writing in that style cannot be ruled out; it also found attacker-side AI recording planned steps as done."
  }
 ],
 "relations": [
  {
   "type": "supports",
   "target": "attackers-run-intrusions-through-agents",
   "note": "These traces are much of the basis on which reporters judge that intrusions were agent-driven, so that finding shares their inference limits."
  }
 ],
 "statusHistory": [
  {
   "status": "corroborated",
   "on": "2026-10-30",
   "why": "Sysdig, Wiz, TrendAI and Google’s threat intelligence group with Mandiant independently describe behavioral traces they use to identify agent or AI involvement.",
   "event": "gtig-mandiant-artex-pentagi-incident-response-2026",
   "kind": "evidence"
  },
  {
   "status": "corroborated",
   "on": "2026-11-17",
   "why": "Correction: the claim describes investigators’ attribution evidence, not a validated ability to distinguish agents from humans or conventional automation. The source accounts independently report these inferences; no detection accuracy is established.",
   "event": "trendai-three-japan-attack-infrastructures-ai-roles-2026",
   "kind": "correction"
  }
 ],
 "halfLifeDays": 180,
 "wouldChange": "Evidence that human operators or conventional automation leave the same traces at similar rates, a published error rate for these indicators, or agent-driven intrusions that routinely avoid them.",
 "fideQuestions": [],
 "methods": [
  "agent-orchestrated-intrusion"
 ],
 "review": "assistant-drafted",
 "addedOn": "2026-11-01"
}