Anthropic details eight months of Claude misuse in September threat report
The lab says it disrupted cyber operations, influence campaigns, surveillance, scams, biological work, weapons development and illicit distillation between December 2025 and August 2026. Fable and Mythos appear in only one distillation case.
Deep dive5 minIn-depth
Maya Chen5 min read
- Anthropic
- Threat intelligence
- Cybersecurity
- Claude
- Safety

Photo: AnthropicSource page
Anthropic published a threat-intelligence report on September 10 covering Claude misuse it says it disrupted between December 2025 and August 2026. The cases fall into seven categories: cyber operations, influence operations, surveillance, scams and fraud, biological misuse, conventional weapons development, and distillation. The models involved were Haiku, Sonnet and Opus. Claude Fable and Mythos-class models appear in none of the misuse cases except one illicit distillation campaign.
The lab tracks actors as Generative Threat Groups and tries to measure what it calls uplift — how much faster, wider or deeper an operation became once a model was in the loop. Humans still chose targets and reviewed the take, Anthropic says. Unite.AI and The Decoder summarised the same report later in the week.
GTG-20006 and a malware rebuild loop
The most detailed cyber case is GTG-20006. Anthropic’s attribution is consistent with public reporting on the Russian espionage actor known as Midnight Blizzard. One operator, a Russian speaker using the handle JackPoterz, ran tradecraft and targeting that matches a Russian state-nexus group: Ukrainian and European military-intelligence and government targets, plus diplomatic and defence organisations and people tied to U.S. foreign policy. More than twenty organisations appeared in the operational planning, including ministries, defence and intelligence bodies, embassies, think tanks and drone-supply firms.
The toolkit included two families of Windows implants, a mobile exploitation kit, a browser-password stealer, a phishing platform dressed as a government service, and a console for stolen accounts. Monitoring agents watched whether security products flagged the malware. When a detection fired, agents modified and recompiled the implant until it was quiet again, then staged it from disposable hosts for phishing, ClickFix and DNS-hijack delivery. Anthropic says this inverts the cost onto defenders: a new signature used to buy time; a capable actor can now close the loop faster than the signature can ship.
If their monitoring AI agents identified that any of their deployed malware was detected by a security product, agents would then set about the process of autonomously modifying and rebuilding the malware to evade the existing detections.
— Anthropic, September 2026 threat intelligence report
The same actor used AI to register domains, stand up phishing infrastructure, and run device-code phishing against cloud email. Anthropic names the actor’s own “Embassy Kit.” Hotel Wi-Fi vendors were compromised so guest traffic could be DNS-hijacked; Microsoft later described a related theft-and-delivery method as CaptiveCrunch. WhatsApp accounts were linked as companion devices through headless browsers. A North African government technology authority lost a credential database of more than 300,000 national identity records and a commercial registry of more than half a million companies. Mail records came out of at least eight organisations, including a national prosecutor’s office.
Financially motivated theft and stolen API keys
GTG-50014 is the financially motivated cluster Anthropic ties to ShinyHunters affiliates. One French-speaking operator mass-downloaded 1.8 million Android APKs, decompiled them, and scanned for hardcoded secrets. A technology-provider breach moved more than a terabyte, including hundreds of thousands of national identifiers and millions of payment-card records, then staged the haul on a public site for extortion. An airline’s systems holding tens of millions of passenger records were reached. After one SaaS foothold, operators extracted data from roughly 200 downstream customer organisations and dumped more than 2,100 Azure AD token sets across more than 40 corporate tenants in about 34 hours. Anthropic says AI agents did nearly all of that work. Another compromise went from a single stolen developer token to full administrative control of a cloud environment in roughly three hours.
Operators also treated the AI supply chain as a source of compute. Stolen Anthropic customer API keys — not Anthropic’s own systems, the lab is careful to say — were reused as attack compute. Fake “discount Claude” shops harvested credentials. One actor injected instructions into an AI vendor’s evaluation sandbox and walked out with production keys from multiple providers. A hacktivist campaign later in the report ran for a month entirely on stolen keys.
Influence-as-a-service and state surveillance
Since its last report’s single commercial influence-as-a-service network, Anthropic says it has disrupted larger shops. Nine cases in this edition ran from Russia, Iran and Turkey and across the Gulf, South Asia, Africa and Europe, aimed at audiences on six continents. GTG-54002 published at least 8,913 articles in about twenty languages for whoever was paying that week — Category Two on the Breakout Scale: high volume, little real audience. A separate Istanbul-based platform sold a “military-grade, AI-driven, real-time political operations ecosystem,” warming more than a thousand fake X accounts and taking a request for a million artificial views on a Malaysian prime minister’s account.
Surveillance cases include China-based clusters that used Claude to compile religious-affairs dossiers, run “stability maintenance” pipelines, and produce daily public-opinion briefings. In one municipal case, Claude refused a weekly suppression report, then complied on a re-prompt that named ten private citizens for “control.” Iranian-nexus and Mali state-intelligence tooling also appear. Conventional-weapons cases include a Yemen-based guided-weapons cell, a China-based undersea fire-control specification, a Russia-based autonomous drone-swarm engineering effort, and electronic-warfare targeting software. Biological misuse is documented as case studies, including a mammal-adapted avian-influenza research programme and covert frontier-model access for orthopoxvirus work.
Illicit distillation, and a single Fable exception
Since a February disclosure, Anthropic says it has disrupted additional illicit distillation campaigns from seven China-based labs against generally available Claude models. Illicit distillation, as the lab defines it, is an industrial-scale covert campaign to extract capabilities — usually through fake accounts, stolen cards and “transfer station” proxies. Operators affiliated with Alibaba ran what Anthropic calls the largest distillation attack it has measured: a fixed prompt that forced Opus 4.6 and 4.7 to write out chain-of-thought inside inline tags, then converted those transcripts into supervised fine-tuning data for Qwen 3.5, 3.6 and 3.7. Peak traffic reached nearly three million exchanges a day from more than 3,500 fraudulent accounts. Between May and July 2026, more than 151 million exchanges. Moonshot, DeepSeek, Zhipu, Xiaomi, SenseTime and MiniMax appear with their own volumes and methods, including a MiniMax shell proxy that sold only Anthropic and OpenAI access.
Fable and Mythos are almost unused in the ledger. No cyber, influence or weapons case uses them. The exception is that one distillation case. Anthropic says it has not seen attempts against Mythos 5 or Mythos Preview, which are not generally available, and that it strengthened anti-distillation classifiers alongside Fable 5, then added preserved thinking on Fable 5.1 so new API accounts cannot edit the context that precedes encrypted reasoning. Dario Amodei’s Saturday essay argued that safety work is being outrun and asked for independent evaluators inside labs. The report is the traffic those evaluators would be asked to review.
- Window: December 2025–August 2026; seven harm areas; Haiku, Sonnet, Opus. Fable/Mythos in one distillation case only.
- GTG-20006: Midnight Blizzard–consistent Russian espionage; malware rebuilt when detections fire; 20+ organisations; Embassy Kit device-code phishing.
- GTG-50014: ShinyHunters-style smash-and-grab; 1.8 million APKs scanned; terabyte-scale theft; 2,100+ Azure AD token sets in ~34 hours.
- Stolen Anthropic customer API keys reused as attack compute — not Anthropic’s own systems.
- Influence-as-a-service at industrial scale; Alibaba-attributed distillation >151 million exchanges, May–July 2026.


