Hackers are using AI agents to breach companies in just hours, Anthropic says

Hackers are increasingly handing over large parts of cyberattacks to AI agents, allowing breaches that once required teams of skilled operators to be completed in just hours, according to Anthropic.

The AI company says its latest threat-intelligence investigation found malicious actors using autonomous AI systems for reconnaissance, exploitation, credential harvesting and data theft — sometimes against multiple victims simultaneously.

The result is a dramatic acceleration in attack speed and scale.

Anthropic said it observed breaches completed in two to three hours, while individual operators handled dozens of victims in parallel.

In one case, an attacker went from a single stolen developer token to full administrative control of a victim’s cloud environment in roughly three hours.

Another breach of an enterprise software company progressed from initial access to bulk data theft in only hours.

AI agents do much of the work

The attacks themselves don't necessarily use previously unknown hacking techniques.

Anthropic said many relied on familiar methods such as stolen credentials, unpatched devices, exposed online services, SQL injection and phishing.

What has changed is how much of the work can now be delegated to AI.

Threat actors are using AI models to perform tasks including reconnaissance, finding vulnerabilities, writing tools, running commands against victim networks, collecting credentials and analysing stolen data.

Some operations used multiple AI agents simultaneously.

A lead agent could break a job into smaller tasks and dispatch them to numerous AI subagents working in parallel.

In the most autonomous cases identified by Anthropic, these systems conducted reconnaissance, exploitation and theft against several victims for hours or days with minimal human supervision.

One hacker can do the work of a team

That could change the economics of cybercrime.

Tasks that previously required teams of skilled operators can increasingly be performed by AI models running continuously and at machine speed.

Anthropic said both criminal groups and state-linked actors are adopting similar techniques.

A financially motivated hacker, hacktivist or espionage operator can use AI to build tools, execute intrusions and process stolen information at volumes that would be difficult for an individual to handle manually.

Massive amounts of data stolen

Some of the compromises Anthropic uncovered were substantial.

In one case involving a technology provider, attackers exfiltrated more than a terabyte of data, including hundreds of thousands of national identifiers and millions of payment-card records.

At an airline, threat actors accessed systems containing tens of millions of passenger records, Anthropic said.

Another supply-chain attack compromised a software provider and let attackers extract data from roughly 200 downstream customer organisations.

Humans are still choosing the targets

Anthropic cautioned that this does not mean AI has completely taken over cyberattacks.

Humans remain heavily involved in key decisions, including choosing targets, deciding how to monetise stolen information, and reviewing the results of operations.

Greater AI autonomy also does not automatically make an attack more damaging.

What AI changes most dramatically is speed, scale and cost.

By automating labour-intensive parts of an attack, AI can lower the level of skill and workforce needed to target an organisation.

Anthropic said that shift could make targets that previously weren’t worth the effort more attractive to cybercriminals.

The company said it banned accounts linked to the malicious operations it discovered and introduced additional measures to detect and disrupt similar activity.

But its broader warning is clear: AI-enabled cyber operations are no longer limited to the most sophisticated attackers, and capabilities once requiring teams of hackers are spreading much more widely.