{
	"id": "52152248-e3a5-4ec3-96da-dad4a936dc25",
	"created_at": "2026-07-28T02:02:47.835568Z",
	"updated_at": "2026-10-20T09:13:27.446731Z",
	"deleted_at": null,
	"sha1_hash": "7f33449f4c10ddf31660abbc251815c21ee997d7",
	"title": "JadePuffer Becomes the First Ransomware Campaign to Let an AI Agent Run the Entire Attack | SecureReading JadePuffer AI Ransomware Uses Autonomous LLM Agent to Automate Entire Cyberattack",
	"llm_title": "",
	"authors": "",
	"file_creation_date": "0001-01-01T00:00:00Z",
	"file_modification_date": "0001-01-01T00:00:00Z",
	"file_size": 1726806,
	"plain_text": "Home\n\nTrainingTraining\n\nHOT NEWS\n\n+\n\n+\n\nNews Topics\n\nSR InfoSR Info\n\nRITICAL\nCADOBE…\n\nVER 1,000\nOPIRACY…\n\nISA\n+ CWarns of…\n\nS Sanctions\n+ UVPN and…\n\nS and Allies\n+ UWarn Russi…\n\nhttps://securereading.com/jadepuffer-ai-ransomware-autonomous-llm-attack/\n\nResources Dashboard Expert Corner\n\nSearch … Search\n\nJadePuffer Becomes the\n\nFirst Ransomware\n\nCampaign to Let an AI\n\nAgent Run the Entire\n\nAttack\n\n 6 JULY 26\n\nA newly discovered ransomware\n\noperation used an autonomous AI agent\n\nto exploit vulnerabilities, adapt to\n\ndefenses, and encrypt critical systems\n\nwith minimal human involvement.\n\nPage 1 of 7\n\nPublic Spotlight\n\nRECENT\nPOSTS\n\nEW\n+ NMALVERTISI…\n\nOTEL WI-\n+ HFI DNS…\n\nhick-fil-A\n+ CConfirm…\n\nEW\n+ N“DOLPHI…\n\nAKE\n+ FCLAUDE…\n\nLATEST\nTHREAT\nNEWS\n\nAKE\n+ FCLAUDE…\n\nRITICAL\n+ CADOBE…\n\nISA\n+ CWarns of…\n\nS Sanctions\n+ UVPN and…\n\nS and Allies\n+ UWarn Russi…\n\nhttps://securereading.com/jadepuffer-ai-ransomware-autonomous-llm-attack/\n\nAI Takes Center Stage\n\nin a New Era of\n\nRansomware\n\nCybercriminals continue to push the\n\nboundaries of automation, and the latest\n\nJadePuffer ransomware campaign marks a\n\nsignificant milestone. Security researchers\n\nhave identified what they believe is the first\n\ndocumented ransomware attack conducted\n\nalmost entirely by an autonomous large\n\nlanguage model (LLM) agent. Instead of relying\n\non human operators to guide each phase of\n\nthe intrusion, the AI agent independently\n\nperformed reconnaissance, credential theft,\n\nPage 2 of 7\n\nPUBLIC\nSPOTLIGHT\n\neta’s\n+ MNew AI…\n\noogle’s IP-\n+ GBased Ad…\n\nrave\n+ BLaunche…\n\npple\n+ ABlocks…\n\npple\n+ AAccount…\n\nhttps://securereading.com/jadepuffer-ai-ransomware-autonomous-llm-attack/\n\nprivilege escalation, persistence, lateral\n\nmovement, and data encryption.\n\nThis development signals a major shift in how\n\nransomware operations may evolve in the\n\ncoming years. While attackers have long used\n\nautomation for repetitive tasks, JadePuffer\n\ndemonstrates that AI can now make decisions,\n\nadapt to unexpected situations, and continue\n\nan attack with minimal human intervention.\n\nInitial Access Through\n\na Known Vulnerability\n\nThe attack began by exploiting CVE-2025-\n\n3248, a remote code execution vulnerability\n\naffecting Langflow, an open-source platform\n\nwidely used for developing LLM-powered\n\napplications. Although the vulnerability had\n\nalready been patched, many internet-facing\n\ndeployments remained exposed due to\n\ndelayed updates and insufficient hardening.\n\nAfter gaining access, the AI agent immediately\n\nbegan gathering intelligence. It extracted\n\nLangflow’s PostgreSQL database, searched for\n\nsensitive environment variables, collected\n\nstored credentials, and identified cloud-related\n\nassets. Additionally, it explored connected\n\nobject storage systems to locate valuable data\n\nthat could support the next stages of the\n\nintrusion.\n\nAI Adapted Like a\n\nHuman Attacker\n\nPage 3 of 7\n\nhttps://securereading.com/jadepuffer-ai-ransomware-autonomous-llm-attack/\n\nOne of the most remarkable aspects of the\n\ncampaign was the AI agent’s ability to adapt\n\nduring the attack. Rather than executing a\n\nfixed sequence of commands, it analyzed\n\nfailures and modified its approach in real\n\ntime.\n\nFor example, when an API request returned\n\nXML instead of the expected JSON format, the\n\nagent automatically adjusted its parsing logic\n\nand successfully continued its reconnaissance.\n\nLikewise, after an authentication attempt\n\nfailed, it quickly refined its strategy and\n\nachieved a successful login within seconds.\n\nThis ability to learn from immediate feedback\n\nmakes AI-driven attacks significantly more\n\nresilient than traditional scripted malware,\n\nwhich often fails when unexpected conditions\n\narise.\n\nPersistence and\n\nLateral Movement\n\nAfter completing reconnaissance, JadePuffer\n\nestablished persistence by creating a\n\nscheduled cron job that regularly\n\ncommunicated with attacker-controlled\n\ninfrastructure. This ensured continued access\n\neven if the initial entry point was disrupted.\n\nThe AI agent then moved laterally into a\n\nproduction server running Alibaba Nacos, a\n\nwidely deployed configuration management\n\nplatform. Using privileged credentials, whose\n\nsource remains unclear, it launched additional\n\nexploitation attempts, including abuse of CVE-\n\nPage 4 of 7\n\nhttps://securereading.com/jadepuffer-ai-ransomware-autonomous-llm-attack/\n\n2021-29441, an authentication bypass\n\nvulnerability capable of creating unauthorized\n\nadministrator accounts.\n\nMeanwhile, the agent also explored potential\n\ncontainer escape techniques, demonstrating\n\nthat it continuously searched for opportunities\n\nto expand its access across the environment.\n\nAutomated\n\nEncryption and\n\nExtortion\n\nThe final stage of the attack focused on data\n\nencryption. JadePuffer encrypted more than\n\n1,300 configuration records stored within\n\nNacos using MySQL encryption functions\n\nbefore deleting the original data. It then\n\ncreated a ransom table containing payment\n\ninstructions, a Bitcoin wallet address, and\n\nattacker contact information.\n\nInterestingly, researchers observed several\n\nindicators suggesting that the AI itself\n\ngenerated portions of the ransomware code.\n\nThese included detailed natural-language\n\ncomments explaining operational decisions\n\nand even the inclusion of a well-known\n\nexample Bitcoin address commonly found in\n\npublic documentation, likely reproduced from\n\nthe AI model’s training data.\n\nResearchers also noted that the encryption\n\nimplementation appeared weaker than the\n\nransom note claimed. While the attackers\n\nreferenced AES-256 encryption, technical\n\nanalysis suggested the actual implementation\n\nPage 5 of 7\n\nhttps://securereading.com/jadepuffer-ai-ransomware-autonomous-llm-attack/\n\nlikely relied on AES-128 operating in ECB\n\nmode. Furthermore, the encryption key\n\nappeared to be randomly generated without\n\nbeing preserved, potentially making recovery\n\nimpossible even if victims decided to pay the\n\nransom.\n\nWhy This Matters for\n\nBusinesses\n\nJadePuffer represents more than just another\n\nransomware family. It demonstrates that\n\nautonomous AI agents are becoming capable\n\nof executing complex cyberattacks that\n\npreviously required experienced human\n\noperators.\n\nFor organizations across the UAE and globally,\n\nthis evolution raises several important\n\nconcerns. AI-assisted attacks can progress\n\nfaster, adapt to changing environments, and\n\nreduce the technical expertise required by\n\ncybercriminals. Consequently, organizations\n\nshould prioritize rapid vulnerability\n\nmanagement, continuous monitoring,\n\nprivileged access protection, and behavioral\n\ndetection technologies capable of identifying\n\nunusual AI-driven activity rather than relying\n\nsolely on signature-based defenses.\n\nAt the same time, AI-powered attacks may also\n\nleave distinctive behavioral patterns that\n\ndefenders can leverage for detection. The\n\nconsistent reasoning style, rapid decision-\n\nmaking, and automated code generation\n\nexhibited by AI agents could provide new\n\nPage 6 of 7\n\nhttps://securereading.com/jadepuffer-ai-ransomware-autonomous-llm-attack/\n\nopportunities for advanced security analytics\n\nand threat hunting.\n\nAs artificial intelligence continues to mature,\n\norganizations should prepare for a future\n\nwhere attackers increasingly rely on\n\nautonomous systems to conduct sophisticated\n\nintrusions at scale. Proactive security controls,\n\ncontinuous visibility, and resilient incident\n\nresponse capabilities will become essential in\n\ndefending against this next generation of\n\ncyber threats.\n\nSHARE:   TWITTER  FACEBOOK  PINTEREST\n\n LINKEDIN\n\nToddyCat’s Umbrij Malware Hijacks Gmail Access\n\nUsing OAuth Tokens Instead of Passwords →\n\n← QuimaRAT Emerges as a Cross-Platform Java\n\nMalware Service Targeting Windows, Linux, and\n\nmacOS\n\nCopyright © 2026 SecureReading\n\nPage 7 of 7\n\nCampaign to Let an AI \nAgent Run the Entire \nAttack \n 6 JULY 26 \nA newly discovered ransomware \noperation used an autonomous AI agent \nto exploit vulnerabilities, adapt to \ndefenses, and encrypt critical systems \nwith minimal human involvement. \n\nRECENT \nPOSTS \nEW+NMALVERTISI… \nOTEL \nWI-+HFIDNS… \nhick-fil-A+CConfirm… \nEW+N“DOLPHI… \nAKE+FCLAUDE… \n\nLA TEST  PUBLIC\nTHREA T SPOTLIGHT\nNEWS  \n\n eta’s\nAKE + M\n New AI…\n+ \nF CLAUDE… \n\n oogle’s IP-\n + G\nRITICAL + Based Ad…\nC \nADOBE… \n\n rave\n +\nISA B Launche…\n+ C \nWarns of… \n\n pple\n + A\nS Sanctions Blocks…\n+ U \nVPN and… \n\n pple\n + A\nS and Allies Account…\n+ \nU Warn Russi…",
	"extraction_quality": 1,
	"language": "EN",
	"sources": [
		"MISPGALAXY",
		"Malpedia"
	],
	"origins": [
		"web"
	],
	"references": [
		"https://securereading.com/jadepuffer-ai-ransomware-autonomous-llm-attack/"
	],
	"report_names": [
		"jadepuffer-ai-ransomware-autonomous-llm-attack"
	],
	"threat_actors": [
		{
			"id": "eb3f4e4d-2573-494d-9739-1be5141cf7b2",
			"created_at": "2022-10-25T16:07:24.471018Z",
			"updated_at": "2026-10-20T02:00:18.294955Z",
			"deleted_at": null,
			"main_name": "Cron",
			"aliases": [],
			"source_name": "ETDA:Cron",
			"tools": [
				"Catelites",
				"Catelites Bot",
				"CronBot",
				"TinyZBot"
			],
			"source_id": "ETDA",
			"reports": null
		},
		{
			"id": "d67df52c-a901-4d55-b287-321818500789",
			"created_at": "2024-04-24T02:00:49.591518Z",
			"updated_at": "2026-10-20T02:00:10.231856Z",
			"deleted_at": null,
			"main_name": "ToddyCat",
			"aliases": [
				"ToddyCat"
			],
			"source_name": "MITRE:ToddyCat",
			"tools": [
				"Cobalt Strike",
				"LoFiSe",
				"China Chopper",
				"netstat",
				"Pcexter",
				"Samurai"
			],
			"source_id": "MITRE",
			"reports": null
		},
		{
			"id": "4c4e1108-8c11-48e3-91e3-95c24042f3a5",
			"created_at": "2022-10-25T16:07:24.329539Z",
			"updated_at": "2026-10-20T02:00:17.227059Z",
			"deleted_at": null,
			"main_name": "ToddyCat",
			"aliases": [
				"Operation Stayin’ Alive",
				"Storm-0247"
			],
			"source_name": "ETDA:ToddyCat",
			"tools": [
				"CHINACHOPPER",
				"China Chopper",
				"Cuthead",
				"FRP",
				"Fast Reverse Proxy",
				"Impacket",
				"Krong",
				"LoFiSe",
				"Ngrok",
				"PcExter",
				"PsExec",
				"SIMPOBOXSPY",
				"Samurai",
				"SinoChopper",
				"SoftEther VPN",
				"TomBerBil",
				"WAExp"
			],
			"source_id": "ETDA",
			"reports": null
		},
		{
			"id": "60d96824-1767-4b97-a6c7-7e9527458007",
			"created_at": "2023-01-06T13:46:39.378701Z",
			"updated_at": "2026-10-20T02:00:15.326357Z",
			"deleted_at": null,
			"main_name": "ToddyCat",
			"aliases": [
				"Websiic"
			],
			"source_name": "MISPGALAXY:ToddyCat",
			"tools": [],
			"source_id": "MISPGALAXY",
			"reports": null
		},
		{
			"id": "4f602ece-d87f-4307-9181-dde821572c11",
			"created_at": "2026-07-28T02:00:03.854946Z",
			"updated_at": "2026-10-20T02:00:23.838919Z",
			"deleted_at": null,
			"main_name": "JadePuffer",
			"aliases": [
				"Storm-3168",
				"JADEPUFFER"
			],
			"source_name": "MISPGALAXY:JadePuffer",
			"tools": [],
			"source_id": "MISPGALAXY",
			"reports": null
		},
		{
			"id": "3fff98c9-ad02-401d-9d4b-f78b5b634f31",
			"created_at": "2023-01-06T13:46:38.376868Z",
			"updated_at": "2026-10-20T02:00:05.274967Z",
			"deleted_at": null,
			"main_name": "Cleaver",
			"aliases": [
				"Alibaba",
				"TG-2889",
				"Cobalt Gypsy",
				"G0003",
				"Operation Cleaver",
				"Op Cleaver",
				"Tarh Andishan"
			],
			"source_name": "MISPGALAXY:Cleaver",
			"tools": [],
			"source_id": "MISPGALAXY",
			"reports": null
		}
	],
	"ts_created_at": 1785204167,
	"ts_updated_at": 1792487607,
	"ts_creation_date": 0,
	"ts_modification_date": 0,
	"files": {
		"pdf": "https://archive.orkl.eu/7f33449f4c10ddf31660abbc251815c21ee997d7.pdf",
		"text": "https://archive.orkl.eu/7f33449f4c10ddf31660abbc251815c21ee997d7.txt",
		"img": "https://archive.orkl.eu/7f33449f4c10ddf31660abbc251815c21ee997d7.jpg"
	}
}