Anthropic reported that between December 2025 and August 2026 it identified several groups using its Claude artificial intelligence for malicious purposes. The company said the detected abuses ranged from cyber operations and influence campaigns to surveillance, fraud, and attempts related to the development of biological and conventional weapons.
ShinyHunters campaign and credential harvesting
Anthropic specifically linked multiple incidents to the ShinyHunters malicious group. According to reporting by Bleeping Computer, a French-speaking member of the group deployed a credential-harvesting process across ten Amazon Web Services EC2 virtual machines. As part of that operation the actor downloaded 1.8 million Android applications from various app stores to search for secrets and credentials.
The recovered data was streamed in real time to a Telegram group. The same malicious actor also conducted an operation that stole GitHub access credentials. Anthropic said the ShinyHunters attack reached credential takeover in roughly 34 hours, and that the bulk of the work was performed by the AI.
Other incidents and exfiltrations
Anthropic reported that additional attacks tied to the group resulted in more than 1 TB of data exfiltrated from a technology company, and that separate intrusions affected an airline and an energy company.
The company also said some of the malicious activity it observed appears to have links to state-backed actors, with indicators pointing to Russia and China.
Anthropic's response
In each identified case Anthropic revoked the malicious actors' access to Claude, banned the involved profiles, and tightened its security controls. The company said it continues to monitor model usage and is improving detection and prevention mechanisms to limit abuse.
Why this matters
The incidents illustrate that advanced language models can be repurposed for large-scale automated abuse: tasks such as data scraping, credential harvesting and account compromise can be accelerated by AI. Rapid detection and remediation by service providers are essential to limit damage, and Anthropic's actions show how a provider can respond when misuse is detected.
