Anthropic Reports Increasingly Autonomous AI Misuse in Cybercrime and Espionage

· Original Report (PDF) ↗

Report discovered via Awesome Annual Security Reports.

Summary

Anthropic’s September 2026 report describes disrupted Claude misuse spanning espionage, credential theft, fraud and surveillance. It reports growing use of autonomous attack workflows and stolen AI credentials between December 2025 and August 2026.

Key points

  • Anthropic reports that threat actors increasingly delegated reconnaissance, exploitation and data theft to multi-agent workflows, while humans retained important targeting and monetization decisions.
  • GTG-20006, whose attribution Anthropic says is consistent with links to Midnight Blizzard, used AI workflows to rebuild detected malware, automate phishing and support espionage against government and defense targets.
  • Suspected ShinyHunters affiliates harvested exposed credentials at scale; one operator scanned 1.8 million Android APKs using TruffleHog and routed verified secrets to Telegram.
  • Anthropic says GTG-10007 combined parallel intrusion workflows with autonomous firmware analysis and exploit testing, targeting roughly 50 organizations and validating previously unknown vulnerabilities in its lab.
  • GTG-50020 stole production API keys through malicious instructions injected into an AI vendor’s evaluation sandbox. Anthropic says attempts to obtain a pre-release Claude model failed and its own systems were not compromised.
  • According to Anthropic, lone hacktivist GTG-50029 gained internal access to at least 14 of 42 tracked targets, stole an estimated 12–26 GB of database dumps, and built the fafsearch doxxing platform.
  • The report also documents deceptive influence networks, surveillance and fraud. Anthropic says it disrupted the reported activity, strengthened safeguards, and shared intelligence with authorities and industry partners where appropriate.

Article Details

Publisher
Anthropic
Report Period
2025-12 through 2026-08
Scope
Selected case studies of Claude misuse across cyber operations, influence operations, surveillance, scams and fraud, biological misuse, conventional weapons development, and distillation.
Key Statistics
  • GTG-50014 operator frkoo ran a credential-harvesting pipeline across 10 AWS EC2 workers, downloading and analyzing 1.8 million distinct Android APKs.
  • One suspected ShinyHunters affiliate extracted data from roughly 200 downstream customer organizations and dumped over 2,100 Azure AD token sets spanning more than 40 corporate tenants in about 34 hours.
  • GTG-50029 gained internal access to at least 14 of 42 tracked target entities and exfiltrated an estimated 12–26 GB of database dumps.
  • GTG-50020 attacked roughly 30 AI companies in about four days; its attempts to obtain a pre-release Claude model all failed.
  • Anthropic attributed over 151 million observed exchanges between May and July 2026 to Alibaba distillation activity.
Recommendations
  • Treat AI API keys and session tokens with the same security rigor as production credentials.
  • Treat agent integrations, sandboxes, proxies, and resellers as part of the organizational attack surface.
  • Purchase AI access only through authorized channels rather than unknown intermediaries offering discounted access.

Indicators of compromise

TypeIndicatorContext
DOMAINawstore[.]cloudDomain listed as an IOC for GTG-50021 fraudulent AI resale and credential harvesting.
DOMAINdocs-viewer[.]orgInfrastructure listed as an IOC for GTG-20006 espionage operations.
DOMAINfafwatch[.]xyzGTG-50029 actor-registered domain associated with doxxing activity.
DOMAINfiftystates[.]newsFabricated news outlet in the GTG-54002 commercial influence-as-a-service network.
DOMAINkiro[.]cheapDomain listed as an IOC for GTG-50021 fraudulent AI resale and credential harvesting.
DOMAINmalaysiapulse[.]comActor-controlled fabricated news front in the GTG-84005 election-manipulation operation.
DOMAINms365-device[.]comInfrastructure listed as an IOC for GTG-20006, which operated device code phishing and token theft.
DOMAINms365-live[.]comInfrastructure listed as an IOC for GTG-20006 espionage operations.
DOMAINnaijapulse[.]orgFabricated news outlet in the GTG-54002 commercial influence-as-a-service network.
DOMAINon-pssword[.]comDomain listed among GTG-50014 operational indicators.
DOMAINsignin-1psswoord[.]comDomain listed among GTG-50014 operational indicators.
DOMAINsys-tools[.]cfdDomain listed as an IOC for GTG-50021 fraudulent AI resale and credential harvesting.
DOMAINwa-connect[.]euInfrastructure listed as an IOC for GTG-20006, which also compromised WhatsApp accounts.
HOSTNAMEautoshop[.]policenationale[.]ccActor-operated storefront selling stolen payment-card records and cardholder personal information.
HOSTNAMEfrntrs-analytics-863060591218[.]europe-west1[.]run[.]appCloud Run origin behind GTG-50029 browser-C2 infrastructure.
HOSTNAMEfrntrs-analytics[.]dedyn[.]ioActor-controlled BeEF browser-C2 hostname used by GTG-50029.
HOSTNAMEiymkjuzymkapovrntoxy[.]supabase[.]coSpecific hosted infrastructure identified as a GTG-50021 IOC.
HOSTNAMEpdfviewer2024[.]b-cdn[.]netSpecific hosted hostname identified as a GTG-20006 IOC.
HOSTNAMEsoraki-proxy[.]20245aad98d27b1b1a2f0f103e1d7ee0[.]workers[.]devSpecific hosted infrastructure listed as a GTG-50014 IOC.
HOSTNAMEupdate[.]stuseamandesilt[.]orgHost listed among GTG-20006 operational indicators.
IPV4162[.]128[.]129[.]106GTG-50014 attacker egress IP observed from 2026-02-20 through 2026-03-10.
IPV4178[.]16[.]54[.]141GTG-50020 attacker egress IP observed from 2026-05-21 through 2026-06-16.
IPV423[.]88[.]118[.]216IP identified as infrastructure supporting the GTG-84005 influence platform.
IPV434[.]156[.]199[.]132GTG-50029 exfiltration endpoint in a hijacked GCP project.
IPV434[.]156[.]95[.]176GTG-50029 exfiltration endpoint in a hijacked GCP project.
IPV437[.]27[.]103[.]22GTG-50020 attacker egress IP observed from 2026-05-26 through 2026-06-13.
IPV62a04:cec0:1185:34f2:a150:7081:caed:448eGTG-50014 attacker egress address observed from 2026-04-06 through 2026-04-07.
SHA256918fa52ae45ed60ba7cc8bdc99c3cbe9ab92e0375ec31fc05d0d4513be11c593Artifact hash listed in the GTG-20006 malware indicators.
SHA256be99857449d2856dd5a84e21c8a3d5e0e01456adb44062ddec5a6b4970d8d42cArtifact hash listed in the GTG-20006 malware indicators.
URLhxxps[:]//s3[.]eu-central-1[.]s4[.]mega[.]io/fuckyoubasil/Specific cloud-storage resource identified as a GTG-50014 exfiltration location.

MITRE ATT&CK

T1078 · Valid AccountsMultiple actors used stolen or exposed credentials to access victim systems and cloud environments.T1098.001 · Additional Cloud CredentialsGTG-50014 operators created cloud API keys in victim accounts to preserve access after credential rotation.T1098.005 · Device RegistrationGTG-20006 automated registration of actor-controlled devices into victim organizational tenants to maintain access.T1114.002 · Remote Email CollectionGTG-20006 bulk-exported compromised mailboxes, including those of drone component manufacturers.T1190 · Exploit Public-Facing ApplicationActors exploited public-facing applications; GTG-50029 successfully exploited a WordPress re-installation race condition on at least four websites.T1505.003 · Web ShellGTG-50029 implanted a webshell hidden among font assets on a compromised web server.T1537 · Transfer Data to Cloud AccountGTG-50029 used exfiltration endpoints in hijacked GCP projects to receive stolen material.T1539 · Steal Web Session CookieGTG-50020 captured verified sessions through reverse-proxy KYC phishing; other actors stole authenticated session tokens.T1555.003 · Credentials from Web BrowsersGTG-20006 used a credential-stealing tool targeting browser password stores.T1562.001 · Disable or Modify ToolsGTG-20006 payloads froze security updates to prevent retrieval and execution of new malware detection signatures.T1567.002 · Exfiltration to Cloud StorageGTG-50014 moved stolen material to consumer cloud storage, including explicitly identified exfiltration storage resources.T1583.001 · DomainsGTG-20006 used AI workflows to research and register domains for phishing infrastructure.T1595 · Active ScanningActors conducted broad scanning and fingerprinting of internet-facing email, remote access, and other systems.

People

Threat Actors

Malware

Vendors

Products

Tools

Countries

Industries

DefenseThey ran operations attacking military intelligence targets in Ukrainian and European governments, as well as diplomatic and defense organizations and individuals connected to US foreign policy.Educationof mail records from at least eight organizations including a national prosecutor office, a military education institute, and a regional intergovernmental organization.EnergyAt an energy company, the operators claimed that they could remotely control the charging current of electric-vehicle chargers installed in customers’ homes.financeThe actor targeted roughly fifty organizations, spanning education, retail, energy, technology, healthcare, finance, manufacturing, as well as multiple government agencies globally.Governmenttool that targets browser password stores, a phishing platform designed to mimic priority targets like government organizations, and an administrative console used to manage compromised accounts.HealthcareThe actor targeted roughly fifty organizations, spanning education, retail, energy, technology, healthcare, finance, manufacturing, as well as multiple government agencies globally.HospitalityNot all targets were direct: to reach their targets indirectly, the actor compromised at least three hospitality vendors that operate hotel guest WiFi.ManufacturingThe actor targeted roughly fifty organizations, spanning education, retail, energy, technology, healthcare, finance, manufacturing, as well as multiple government agencies globally.Retailweeks to conduct secondary attacks, which targeted other organizations including compromising a French retail chain and probing a Web3 identity platform.TechnologyA common theme of the targeting was Ukraine and military drone technology providers and supply chains.

Related Articles