Anthropic Stops World’s First AI-Orchestrated Cyber ​​Espionage Attack

The world’s fastest-changing technology, Artificial Intelligence (AI), is no longer just doing good; it is doing great. Recently, a major cyberespionage attack was carried out using AI, dubbed the first “AI-Orchestrated Cyberattack” in history.

And shockingly, this attack was carried out using Anthropic’s own Cloud AI model.

In mid-September 2025, Anthropic spotted some weird activity. After digging in, it turned out to be a highly advanced espionage campaign. The attackers weren’t just using AI behind the scenes; they let the AI run the show, taking on tasks and carrying out attacks directly.

Anthropic is confident that this was a Chinese-sponsored group. They used Anthropic’s cloud code tools to attempt to breach approximately thirty different locations around the world, and in some cases, succeeded. As far as Anthropic knows, this is the first major cyberattack in which AI performed the majority of the work, without any human involvement.

As soon as they learned of this, they immediately began working to determine the extent of the breach. Over the next ten days, they assessed the damage, shutting down accounts as soon as they found them, alerting affected companies, and staying in constant contact with authorities as they learned more information.

This entire development has changed the game. Now that AI agents can work alone for long periods of time, handling complex tasks without any human assistance, they are a huge asset to productivity. But in the wrong hands, they make it very easy to carry out major cyber attacks.

These attacks are becoming more frequent. To combat them, Anthropic has enhanced its detection tools and developed more robust systems to detect malicious activity. They are constantly working on new ways to detect and investigate such large and widespread attacks.

Anthropic says they’re sharing this information now to help anyone in technology, government, or research who wants to strengthen their security. And they’ll continue to provide updates like this, sharing information about what they’ve found and what’s coming next.

The attack took advantage of AI features that didn’t exist or weren’t very good a year ago.

What happened:-

• Anthropic detected “suspicious activity” in mid-September 2025, which was later determined to be a high-level espionage campaign.
• In this attack, manual hackers used the Cloud Code tool, but surprisingly, the AI ​​itself did 80-90% of the work.
• Anthropic estimates that a Chinese-backed (state-sponsored) hacker group was behind this campaign.
• They targeted approximately 30 global targets (tech companies, financial institutions, chemical manufacturing companies, government agencies).

Anthropic’s report states that the attack consisted of the following steps:

Jailbreaking: The hackers lured Claude under the guise of “security testing” or “defense testing.” They first asked seemingly innocent questions, hoping to gradually break down his guardrails.

Reconnaissance: Claude understood the target’s system structure and focused on the most critical databases. He also discovered vulnerabilities.

Exploit Code Generation: Once vulnerabilities were identified, Claude wrote the exploit code himself.

Credential Harvesting and Data Extraction: Claude extracted usernames and passwords, created backdoors to high-privileged accounts, and stole data.

Documentation: Finally, Claude documented the entire attack – including which credentials were stolen, system analysis, and further steps.

According to Anthropic, humans only intervened in 4-6 major decisions during the entire attack. Claude handled everything else automatically.

This news from Anthropic is a warning – AI is no longer just a tool, but can become an independent agent that, in the wrong hands, could lead to large-scale cyberattacks. But at the same time, the same AI can also help us combat these challenges.

This development points to a new direction for cybersecurity – a world where AI agents will not only be useful for their own sake and capabilities, but also with security at the forefront.

Frequently Asked Questions (FAQs)