Cybercrime has always required patience: reconnaissance, careful probing, days or weeks of quiet groundwork before a breach pays off. That calculus appears to be changing fast, according to a new threat-intelligence report from Anthropic, the company behind the Claude chatbot.

The company’s findings, published in September 2026, describe attackers no longer just consulting automated assistants for tips they are handing over entire stages of an attack to autonomous software that can plan, execute, and adapt on its own. Cybercriminals and state-backed hackers were increasingly using these systems not just to assist with tasks but to orchestrate and execute large portions of cyberattacks, the company said.

From Assistant to Operator

For years, the fear around misuse of smart software centered on it giving bad actors better instructions — a faster way to look something up, a cleaner phishing email. What Anthropic is now describing is a different category of risk entirely. Every layer of offensive operations has reportedly been uplifted, from reconnaissance and tool development to data processing and exploitation, giving attackers a wider and deeper base of knowledge that speeds up both how fast they build new capabilities and how fast they use them.

The practical effect shows up in the timelines. One breach of an enterprise software company took only hours from first access to bulk data theft. Another compromise escalated from a single stolen developer credential to full administrative control of a victim’s cloud environment in roughly three hours, followed by systematic scraping of internal data stores and, in supply-chain cases, of the end customer’s data too.

This isn’t isolated to one skilled hacker with unusual talent. The report finds that large language systems are increasingly embedded into autonomous, multi-agent frameworks that can execute complex tasks at machine speed, reducing the labor and tooling gap between major nation-states and lower-resource actors. In other words, the specialized teams that only well-funded governments could once assemble are now within reach of far smaller, far less sophisticated groups because the heavy lifting is delegated to software instead of staff.

Operators Managing Dozens of Victims at Once

Beyond speed, the report points to scale. One attacker profile internally tracked by Anthropic as a Generative Threat Group was found scraping public vulnerability-disclosure platforms like HackerOne as a reconnaissance step before focused strikes on chosen targets. That same actor kept up a steady operational rhythm, hopping between many compromised organizations with something close to production-line efficiency rather than the artisanal, one-target-at-a-time approach that once defined serious intrusions.

Anthropic’s broader findings, drawn from an eight-month window between December 2025 and August 2026, span seven distinct harm categories: cyber operations, surveillance, influence campaigns, conventional weapons development, biological misuse, scams and fraud, and illicit model distillation. The company said the incidents it chose to highlight were selected because they were notable or technically novel not because they represent a full accounting of everything abusive happening on its platform.

Among the disclosed cases, Anthropic said it disrupted a suspected Russia-linked espionage campaign and separate attempts by companies based in China to copy Claude’s underlying capabilities. Notable incidents involved Chinese labs and the Russian hacking group Midnight Blizzard, according to the company’s account.

Old Techniques, New Velocity

Importantly, the report doesn’t describe a wave of exotic, never-before-seen hacking methods. The building blocks — credential theft, privilege escalation, lateral movement through a network remain largely familiar to security professionals. What’s changed is who can execute them, and how quickly. Tasks that once required a coordinated team working in shifts can now run largely unattended, with a human operator stepping in mainly to pick a target and review the results once the job is mostly done.

That shift matters for defenders as much as attackers. Security researchers have noted that conventional detection playbooks, built around the assumption that sophistication signals a well-resourced adversary, may no longer hold. A lone operator with the right access to capable automated tooling can now move with a speed and reach that used to be the signature of state-level operations meaning organizations can no longer use “this looks too advanced for a small-time criminal” as a reliable filter for how seriously to treat an intrusion.

What Comes Next

Anthropic said it has banned the accounts tied to the disrupted activity, rolled out new detection methods, and shared relevant intelligence with authorities. But the underlying tension the report surfaces isn’t easily solved by any single company’s enforcement action: the same systems that help a legitimate developer write and debug code are, by their nature, equally capable of helping someone find and exploit a flaw in somebody else’s.

As these tools keep getting more capable and more widely available, the gap between what a nation-state can do and what a single motivated individual can do looks set to keep narrowing putting pressure on security teams everywhere to assume that any attacker, however small, might now be moving at industrial speed.