01 // EVIDENCESeen in real attacks
Sorted by strength of evidence
Every row links its primary source
| Attacker | What it did | ||||||
|---|---|---|---|---|---|---|---|
| 01 | ARTEXmulti-agent pentest platform; repo closed Oct 2026, forks live onReconExploitationPost-exploitation | CONFIRMED | VAPOR PANDAChina-nexus espionageStorm-0940China-state aligned+5 more actorsStorm-3358Cambodia-based, targets financial firmsCRAFTY FALCONtargets Indian financial firmsUnnamed actorChinese-speaking, financially motivatedOnyx SleetNorth Korean state actor (Microsoft)Pearl SleetNorth Korean state actor (Microsoft) |
One controller hit 200+ organizations in 11 days, with access confirmed at nine. Storm-0940 used it to try sign-ins at 2,250+ Microsoft 365 tenants. Zscaler tied a theft of personal data on about 67,000 people from South Korean financial firms to it. Check Point read operator consoles: one operator stole 300+ iOS code-signing certificates. GMO found it on the attack infrastructure behind breaches of Japanese companies. Microsoft names two North Korean actors using it. | 200+organizations targeted | Yes | CrowdStrike ↗ Microsoft ↗ Microsoft II ↗ Zscaler ↗ CERT-In ↗ FBI ↗ GTIG · Mandiant ↗ GTIG Nov ↗ Check Point ↗ Datadog ↗ Unit 42 ARTEX ↗ Amazon ↗ Microsoft III ↗ GMO Cybersecurity ↗ 2026-10 · 11 |
| 02 | Cairnmulti-agent pentest frameworkReconExploitationPost-exploitation | CONFIRMED | CL-STA-1238likely China-nexus espionageCRAFTY FALCONtargets Indian financial firms+3 more actorsRetail card-theft operatorunnamed, financially motivated, prompts in ChineseAIA0001NorForge cluster: Chinese-language, e-commerce accessAIA0002NorForge cluster: English-language, card and crypto theft |
Given a domain and a one-line goal, ran 25 operations against at least 19 organizations, with access or data taken at 10. Gambit counted 105 attack projects launched between 10 and 15 September and at least 27 companies compromised; NorForge found 20+ exposed panels and confirmed a live uploader web shell at a hardware-parts vendor. Google separately saw it used against a US aerospace target. | 27+companies compromised in one campaign | Yes | 2026-09 · 11 |
| 03 | CyberStrikeAILLM orchestrator with a built-in C2ReconScanningVuln discoveryExploitationPost-exploitation | CONFIRMED | Unattributed operatorsKaspersky could not link them to a known groupCRAFTY FALCONtargets Indian financial firms+1 more actorUnattributed operator (Japan)TrendAI: Simplified-Chinese console, targets in Japan |
Implant deployed on education-sector networks in Latin America and Africa. The operator set the plan; the model on the command server chose and issued the technical steps: credential dumping, lateral movement, DCSync and Golden Ticket attacks, then database export. TrendAI separately confirmed it running 44 agents against at least 19 organizations, 16 or more in Japan, in October; deep compromise reached only a small share. Team Cymru found its banner on an IP Amazon had published from an AI-augmented actor's infrastructure, with NetFlow to the FortiGate devices that actor targeted. | 3–4 daysto domain admin | Yes | 2026-03 · 11 |
| 04 | PentestGPTautonomous pentest agentReconScanningVuln discoveryExploitationResearch | CONFIRMED | UAT-10147Chinese-speaking, financially motivatedUnnamed operatorSpanish- and English-speaking | Installed on the attacker’s command server, scanned web servers and broke into at least one site. An operator running a rebuilt fork compromised at least 15 Argentine healthcare organizations. | 170,000URLs targeted | Yes | 2026-08 · 10 |
| 05 | PentAGImulti-agent pentest systemScanningExploitationPost-exploitation | CONFIRMED | GTG-50020Russian-speaking criminalsGTG-50029French-speaking hacktivistCRAFTY FALCONtargets Indian financial firms | Named by Anthropic as a framework both groups used. One stole production AI API keys; the other dumped European political databases. Google saw most PentAGI activity stop at scanning, but some reach intrusion and persistence at multiple organizations. | 12–26 GBdatabases taken | Yes | 2026-09 · 11 |
| 06 | OpenClawgeneral-purpose agentExploitationPost-exploitation | CONFIRMED | Unnamed operatorSpanish-speaking“Recon” operatorfinancially motivated | An agent on OpenClaw exploited a Telegram Mini App, dumped its database of 1,300+ users and verified the app's bot token. Google also found an .openclaw/ directory on a credential-harvesting server holding 23,800+ secrets. | 1,300+user records dumped | Yes | 2026-08 · 09 |
| 07 | Hermes Agentopen-source autonomous agent with memory and self-written skillsReconScanningExploitationPost-exploitation | CONFIRMED | knaithe / KnYuanChinese-speakingRetail card-theft operatorunnamed, financially motivated, prompts in Chinese+2 more actorsSilent LynxCentral Asia espionage cluster, medium-confidence Kazakhstan linkBOSSAI operatorlone operator; Acronis found no clear affiliation |
Four operations. Unit 42: exploit attempts against Langflow and n8n failed and a Citrix exploit run by hand got in. Gambit: the operator's console in a retail card-theft campaign. Acronis: the main attack agent in a lone operator's 12-agent mesh, extracting data at a Thai university. CloudSEK: handled recon, collection and log cleanup for an espionage cluster inside Tajikistan’s railway network. | 4operations reported | Yes | 2026-07 · 11 |
| 08 | oh-my-claudecodeadd-on layer for Claude CodeScanning | CONFIRMED | Zerofot | Ran the IT and DevOps for a credential harvester its operator built with Codex and Claude Code: brought up, restarted and debugged the live scanner and its proxies. | 1,000shosts harvested | Tooling | 2026-08-31 |
| 09 | Strixmulti-agent pentest frameworkReconScanningVuln discoveryExploitationExploit devPost-exploitation | CONFIRMED | Suspected China-linked actorplus an attacker on hijacked AI serversRetail card-theft operatorunnamed, financially motivated, prompts in Chinese | In GTIG's suspected China-linked campaign, validated vulnerabilities. A honeypot caught it aimed at a live French shop with no permission prompts. Gambit saw it run 146 times against 138 hosts as the vulnerability-search stage of a retail card-theft operation, and some of its reports opened the next stage of exploitation, handed to Cairn. | 2,000+unattended steps | Discovery only | 2026-05 · 09 |
| 10 | HexStrike AIMCP server wrapping 150+ offensive toolsReconScanningExploit devResearch | OBSERVED | Suspected China-linked actorthe same actor as the Strix caseBOSSAI operatorlone operator; Acronis found no clear affiliation | Moved between recon tools on its own against a Japanese tech firm and an East Asian security company. Google listed it among frameworks abused in incident response in October, then described what it had seen of HexStrike as experimentation in November. Acronis found it installed as an MCP tool server on a lone operator's attack node and wired into every agent that needed it; the report ties no breach to it. | 2named targets | Not stated | 2026-05 · 11 |
| 11 | DeepAuditmulti-agent source-code auditorReconScanningVuln discoveryExploit dev | OBSERVED | UAT-10147Chinese-speaking, financially motivated | Installed on the attacker's management server, and named among the frameworks the actor used to automate intrusions. Talos did not see it used against victims. | 1attacker server | Not stated | 2026-08-20 |
| 12 | AntiFoxARTEX derivative; Chinese-language agentic pentest platformExploitationPost-exploitation | OBSERVED | Unattributed AntiFox operatorsArctic Wolf names no actor; AntiFox hosted on Alibaba Cloud | Infrastructure hosting AntiFox on Alibaba Cloud was “associated with” confirmed intrusions: a web exploit, an IIS web shell, AWS credential checks and a successful read of a production secret. Arctic Wolf ties the activity to the infrastructure, not to the tool itself. | 1+production secret read | Associated | 2026-11-10 |
| 13 | CyberStrikeopen-source AI pentest agent; not CyberStrikeAIReconVuln discoveryExploitation | OBSERVED | BOSSAI operatorlone operator; Acronis found no clear affiliation | One agent in a lone operator's 12-agent attack mesh. Acronis says it appears to have run vulnerability assessments and exploitation attempts against live targets, but its outputs could not be verified. | 1 of 12agents in the operator's mesh | Not stated | 2026-11-04 |
| 14 | HalberdStrikepentest harness based on PentestGPTReconVuln discovery | OBSERVED | BOSSAI operatorlone operator; Acronis found no clear affiliation | Mapped the attack surface across at least 20 projects, from a Taiwanese emergency-services portal to 50+ commercial sites, and passed its reports to other agents for exploitation. Over half its high and critical findings were false positives or misclassifications. | 20+projects mapped | Not stated | 2026-11-04 |
02 // FIELDNo attack on record
140 more with no attack on record
GitHub facts as of 2026-11-17
03 // SOURCESPrimary reports
32 reports from 22 publishers
Newest first
04 // METHODHow this is built
Evidence reviewed 2026-11-17
GitHub facts 2026-11-17
Tying a harness to an attackA harness enters table 01 only when a security vendor, IR firm, CERT, government agency or AI lab names it in its own report of a real operation, and the row links that report. We never tie a harness to an attack from overlapping TTPs, shared tooling or our own judgment. Forum posts, marketing and news stories that cite no primary report do not count. When a report hedges, the row hedges with it.
Evidence tiersConfirmed means the report shows the harness in a real operation. Observed means it was aimed at real targets or found on attacker infrastructure, with no breach stated. No attack on record means none in primary reporting, which does not prove none happened.
Use tagsOutlined tags come from each project's own README and are its claim, not a tested result. Filled tags come from attack reporting. Hover any tag to read the exact sentence behind it, click to open the README or report, or get every sentence in data.json. A tool table that lists sqlmap does not earn Exploitation; the README has to say the project attacks targets.
What the reports showIn the ARTEX and Cairn cases the frameworks ran long stretches of an intrusion on their own, mostly through old weaknesses: weak passwords, auth bypasses, known CVEs. CyberStrikeAI's agent ran textbook domain-compromise steps for a human planner, and a second operator ran 44 of its agents against organizations in Japan, with deep compromise at only a few. PentestGPT, PentAGI and the OpenClaw Telegram case also exploited live targets on their own. Hermes Agent's exploits failed in July and a human got in; in three later operations it was the attacker's console or main agent, including data extraction at a Thai university. Strix, first seen validating vulnerabilities, was the discovery stage of a retail card-theft campaign, and some of its reports went to Cairn. HexStrike, DeepAudit and HalberdStrike were seen only mapping, scanning or sitting on attacker servers. CyberStrike's outputs could not be verified, and AntiFox is tied to intrusions only through the infrastructure that hosted it.
Scope and rosterThe field table lists open-source projects under an OSI license that put an LLM in the loop of offensive security work, on public GitHub with 100 stars or more, or with a primary report that names them. The attack table needs only a primary report that names an open-source harness, at any star count, and it includes general agents threat actors were seen using. Source-available projects such as CAI are left out, and so are commercial agents like Claude Code, Codex and Gemini CLI.
The incidents ledgerOur sister site, hackingincidents.ai, records the attacks themselves: every cyberattack where AI was known or strongly suspected to be used, graded and sourced. A row here that the ledger mentions links to those incidents under what it did (9 harnesses, 12 incidents, checked 2026-11-17), and the ledger links back to the rows here.
Freshness and the dataGitHub facts and README tag sentences are re-checked on every refresh, and the refresh fails once the evidence review is older than 21 days. Each evidence pass gets an independent model review against its sources. Everything here renders from one dataset: data.json, mirrored with the full research notes in Substrate ↗. Every data source and step is laid out on the About page.