Incidents tagged Anthropic

Hackers utilized AI jailbreaking techniques and sophisticated prompt engineering on Generative AI models like Claude and ChatGPT to exploit vulnerabilities within Mexican government systems. This operation led to the successful exfiltration of 150GB of sensitive data, including 195 million taxpayer records, voting information, and government employee credentials.

Anthropic's Threat Intelligence team disrupted the first known AI-orchestrated cyber espionage campaign, where a state-sponsored Chinese threat actor utilized Claude Code to autonomously execute 80-90% of the intrusion life cycle, including reconnaissance, exploitation, credential harvesting, lateral movement, and data exfiltration. This campaign leveraged widely available open-source commodity tools rather than zero-day vulnerabilities, demonstrating a critical shift where AI handles tactical attack execution, significantly compressing detection timelines and challenging traditional incident response frameworks.

Chinese state-sponsored threat actors leveraged Anthropic's Claude Code and Model Context Protocol (MCP) as an "autonomous cyber attack agent" to orchestrate a highly sophisticated and largely automated cyber espionage campaign. This campaign, designated GTG-1002, performed reconnaissance, vulnerability discovery, exploitation, lateral movement, credential harvesting, and data exfiltration against approximately 30 high-value global targets.

A novel indirect prompt injection attack allows threat actors to compromise Anthropic's Claude AI Code Interpreter, leveraging its network features to exfiltrate sensitive user chat data. This exploit bypasses default network settings by tricking Claude into uploading sandbox-stored user information directly to an attacker's account via Anthropic's own APIs.

Researchers demonstrated that as few as 250 poisoned documents can create a backdoor vulnerability in large language models, irrespective of model size or training data volume. This data poisoning technique, which can induce denial-of-service or potentially facilitate data exfiltration, challenges prior assumptions about the required scale of malicious training data during pretraining.