Anthropic: Autonomous AI Executed Attacks Targeting 70 Organizations Across Two Continents

In its September 2026 security report, Anthropic revealed seven categories of harm resulting from the misuse of artificial intelligence between December 2025 and August 2026. This marks the first official documentation by a major AI company of fully autonomous cyber offensive operations conducted by AI with minimal human supervision.

The report tracked the Russian espionage group GTG-20006, which targeted more than 20 organizations in Ukraine and Europe, while documenting the activity of the Chinese group GTG-10007, which targeted around 50 organizations worldwide. Under reported minimal human supervision, the latter produced over a dozen potential undiscovered security vulnerabilities in just one month.

The report documented a single breach that resulted in the theft of over 300,000 identity records, with the exfiltration of an entire terabyte of data taking only between two and three hours from the initial point of compromise—a figure that redefines the timeframe available for cyber defense teams to intervene before a breach is complete.

On the influence campaign front, Anthropic detected 9 propaganda campaigns spanning 6 continents, which included more than 70 fake news outlets and up to a thousand synthetic accounts per campaign, seeking to influence democratic elections in Moldova, Kenya, and Malaysia. The report confirmed that two separate groups breached AI vendors' evaluation sandbox environments, implying that isolated testing environments are no longer a sufficient guarantee of security.

This means that organizations in the Arab region relying on test environments to evaluate AI models now face direct, previously uncalculated risks, necessitating an urgent review of usage policies and oversight of AI tools within critical infrastructure.

What do these terms mean?

GTG-20006 / GTG-10007: Classification designations used by Anthropic to identify advanced threat groups suspected of state affiliation, similar to classifications used by other security firms such as APT.

Zero-day vulnerability: A security flaw unknown to the developer company, meaning no update or patch is available for it, making it one of the most dangerous tools in a hacker's arsenal.

Evaluation Sandbox Environment: An isolated computing environment used to test software and models safely, separate from core systems—though the report documented two groups breaching them.