Anthropic Report Details AI Agent Use In Cyber Operations

Anthropic’s latest threat report shows attackers using AI agents to coordinate reconnaissance, exploitation and data theft, extending misuse beyond one-off requests for malicious code, Indian Express reported.

The September 2026 misuse review covers cases found by Anthropic’s threat-intelligence staff over the period from December 2025 through August 2026.

It is the company’s fourth public review of Claude abuse and groups the activity into seven areas: cyber campaigns, online influence, monitoring of people, fraud schemes, biological-risk assistance, weapons-related misuse and efforts to copy model capabilities.

Anthropic said it disrupted each activity described in the report, used the findings to strengthen safeguards and shared intelligence with authorities or industry partners where appropriate.

The cases involved Claude Haiku, Sonnet and Opus.

Its most advanced models, Claude Fable and Mythos, were not involved except in one illicit distillation case.

The most important shift is operational.

Earlier misuse often meant a person asked a model to write malware, explain a vulnerability or draft a phishing email, then used that output manually.

In the newer cases, multi-agent systems carried out several stages of an attack while humans set targets and reviewed outcomes.

That change can compress work that previously required a larger team.

Anthropic described state-sponsored groups, financially motivated criminals and individual operators using AI for reconnaissance, tool development, data processing, exfiltration and other tasks.

If AI can make advanced tradecraft easier to repeat, the apparent sophistication of an intrusion may reveal less about the skill of the person behind it.

One case centered on a threat actor Anthropic tracks as GTG-20006.

The company assessed the activity as consistent with public reporting that links the group to Russian-linked Midnight Blizzard.

The campaign reached defence-linked companies, diplomatic bodies, government agencies and military intelligence targets across Ukraine, Europe and selected locations elsewhere.

AI was used across much of that operation, including reconnaissance, phishing infrastructure, persistence in compromised systems, data extraction and malware modification.

The malware case is especially sensitive because agents monitored whether security products detected the tools, then modified and rebuilt them in an attempt to evade those defences.

Scale is another pressure point.

The group studied by Anthropic targeted more than 20 organisations, and one case involved AI helping organise hundreds of gigabytes of stolen data.

The risk is not only a new class of attack, but the cheaper and faster repetition of familiar ones.

The report also moved beyond traditional hacking.

Anthropic identified misuse involving fake dating apps designed to defraud users and surveillance systems used to identify and monitor dissidents.

The broader pattern is that capable models can reduce the labour, technical depth or organisational resources needed for abuse.

For defenders, the response path is becoming more active.

Anthropic argues that security teams will need AI not just to answer incidents, but to find vulnerabilities, detect suspicious activity and harden systems before attackers exploit them.

The report does not claim that cyberattacks have become fully autonomous; humans still choose goals and instructions.

The operational burden, however, is shifting from performing every step to directing systems that can perform many of them at once.