Claude attacks expose rise of autonomous hacking

The latest reports of attackers using Anthropic’s Claude are not simply another example of criminals experimenting with generative AI. They point to something more significant: the emergence of AI agents capable of carrying out parts of an attack chain with minimal human input.

 

Researchers at OALABS recently documented a campaign in which a relatively low-skilled attacker used Claude alongside OpenAI Codex to automate reconnaissance, analyse stolen data and generate code during attacks against multiple organisations. Rather than writing malware from scratch, the operators relied on AI agents to accelerate tasks that would previously have required more technical expertise.

 

The incident reflects a broader shift in cyber-security. For years, AI has been used to generate phishing emails or write snippets of malicious code. Agentic AI goes further by planning tasks, interacting with tools and adapting its behaviour based on the results it receives.

 

Anthropic has itself warned that it has observed threat actors using Claude for increasingly sophisticated cyber activity. In its latest threat intelligence findings, the company said AI is lowering the barrier to entry for offensive cyber operations, while also helping experienced attackers work faster and at greater scale. The report maps real-world misuse across the MITRE ATT&CK framework, showing AI supporting reconnaissance, credential theft, malware development and social engineering.

 

This trend matters because the challenge is no longer just whether AI can generate malicious code. The more pressing question is how much of an attack can be delegated to autonomous agents. Recent academic research suggests that frontier AI models are becoming increasingly capable of completing realistic multi-stage intrusion tasks, even if they are not yet able to independently execute complex attacks end-to-end.

 

For defenders, this shifts the focus from detecting AI-generated malware to recognising AI-enabled workflows. Security teams may face adversaries who can identify vulnerabilities, modify tactics and automate repetitive stages of an intrusion far more quickly than before.

 

While current safeguards still limit what commercial AI models will do, recent events suggest that the conversation is moving beyond individual prompts. As agentic AI becomes more capable, organisations may need to prepare for attackers who use AI not simply as a coding assistant, but as an operational partner.