Skip to content
Unlisted Report logoUnlisted ReportSubscribe
AI Security

Anthropic Report Details How Hackers Misuse AI Models

Anthropic says it disrupted Russian espionage, Chinese model-copying, scam apps and surveillance tools that tried to abuse its Claude models.

Anthropic has published a report describing how threat actors tried to misuse its Claude AI models between December 2025 and August 2026, according to the company.

What the report covers

Anthropic says it disrupted activity across seven areas:

  • Cyber operations
  • Influence operations
  • Surveillance
  • Scams and fraud
  • Biological misuse
  • Conventional weapons development
  • "Distillation" — copying a model's abilities

The actors included suspected state-sponsored groups, criminals, commercial spyware vendors and state propaganda bodies. Cases ranged from a network of fake dating apps built to defraud users to surveillance systems designed to identify and monitor dissidents. Anthropic said the misuse it documented involved its Claude Haiku, Sonnet and Opus models; none of the cases involved its newer Fable or Mythos-class models, with the exception of a single distillation case, according to Anthropic's own report.

The report was published on 10 September 2026 and is the company's fourth such disclosure, following earlier threat reports in March, August and November 2025. In each case, Anthropic says it disrupted the activity, used the findings to strengthen its safeguards, and shared intelligence with authorities and industry partners where appropriate.

Headline cases

  • A suspected Russia-linked espionage campaign, tracked internally by Anthropic as GTG-20006 and assessed as consistent with the Russian state-linked group Midnight Blizzard (also known as APT29 or Cozy Bear), used Claude to automatically rebuild its malware whenever security tools detected it, The Hacker News reports.
  • Anthropic accused Chinese AI firms of trying to extract and replicate Claude's abilities, Reuters reports. Coverage from The Decoder named Alibaba and DeepSeek among the Chinese labs accused of running covert networks to extract training data from Claude at scale, or of secretly rerouting their own customers' requests through Claude, in some cases processing sensitive information such as government surveillance data in the mix, according to The Decoder.
  • Other disrupted activity reportedly included credential-harvesting work associated with the ShinyHunters extortion group, a software-supply-chain intrusion affecting more than 200 downstream organizations, suspected Chinese-linked exploit research, and politically motivated theft and doxxing, according to threat-intelligence aggregator Mallory.
  • Anthropic also documented groups using Claude to help write software for weapons systems, including missiles and autonomous drone swarms, and cases where legitimate and harmful biological-research intent were described as "nearly impossible to tell apart," prompting the company to call for access limits on verified users for the most sensitive biological-research capabilities, according to The Decoder's summary.

The scale of what Anthropic disclosed

The report runs to more than 140 pages, with the largest section devoted to cyber operations. According to coverage from TheNextWeb, the central argument of that chapter is that AI has significantly narrowed the gap between well-resourced state hacking teams and lone individual operators — Anthropic pointed to cases involving a hacktivist using stolen API keys, financially motivated individual criminals, and a state espionage operator, each of which ran multi-victim campaigns that a year earlier would have required a team of specialists, according to TheNextWeb. One security outlet summarized the core finding of the cyber chapter as: "sophisticated attacks no longer require sophisticated attackers, and sophistication is no longer a reliable signal for attribution," according to The Decoder.

Why it matters

AI tools speed up attackers just as they speed up everyone else. Reports like this show defenders what real-world abuse looks like, and they put pressure on AI companies to detect and share misuse. Anthropic frames this publication practice explicitly as a responsibility: "We're publishing this work because we believe we have a responsibility to disclose malicious misuse of our services. As models become increasingly capable, their risks will increase, unless AI developers and society's defenders act to make them safer," the company states in its own report.

The report also marks a shift in how attribution works for defenders. When the technical sophistication of an attack can no longer be assumed to reflect the sophistication of the group behind it, security teams and governments may need new ways to assess the risk posed by a given threat actor, independent of the tools that actor has historically been able to build on its own.

Anthropic's own enforcement response

Anthropic has said that in each disrupted case it banned the accounts involved, used what it learned to tighten technical safeguards against future misuse, and shared relevant threat intelligence — including indicators of compromise — with law enforcement and industry partners where appropriate. The company has also flagged that AI API keys, session tokens and AI-integrated services are themselves becoming valuable targets for theft, since stolen access can be resold, used to obtain victim-funded compute, or used to obscure the true identity of whoever is running an attack, according to Mallory's summary of the disclosed activity.

What organizations should take from it

  • Assume attackers can update their malware quickly; rely on behaviour-based detection, not only known signatures.
  • Treat AI tools your staff use as part of your security policy, including logging and reviewing how internal teams use third-party AI coding or research assistants.
  • Protect API keys and session tokens for AI services with the same rigor you apply to cloud credentials, since they are now a direct target for theft and resale.
  • Build incident response plans that assume an adversary can iterate on tooling in near real time, rather than relying on periodic manual updates to detection rules.

What readers should do

  • If your organization uses AI coding assistants or agent frameworks, review access logs for unusual volumes of automated activity that could indicate account compromise or misuse.
  • Treat any security product or vendor with privileged access to your systems as part of your attack surface, and require prompt patching commitments from vendors.
  • Stay aware that AI-themed news, including threat reports like this one, can itself be used as a hook for phishing — always verify sourcing before clicking links in security-alert emails referencing AI misuse.
  • Follow updates from AI vendors' own threat-intelligence teams, which increasingly publish actionable indicators of compromise that defenders can use directly.

Context: why AI vendors now publish threat reports

Anthropic's quarterly-style disclosures reflect a broader shift among frontier AI labs toward publishing dedicated threat-intelligence research, similar to how cloud providers and antivirus vendors have long published reports on abuse of their own platforms. Because large language models can be used both defensively and offensively, companies like Anthropic argue that transparency about documented misuse — rather than silence — gives the wider security community a chance to build matching defenses before techniques spread further.

By · Published · Updated · 9 min read

Sources & Verification

Additional reporting

Read next