New Delhi: Anthropic has flagged a sharp shift in how cybercriminals are using AI tools, and the warning is not the usual “AI can write bad code” story.
Its new analysis says attackers are now using AI deeper inside cyber operations, including account discovery, credential theft, lateral movement and live network work.
The report studied 832 accounts banned for malicious cyber activity between March 2025 and March 2026. Anthropic said these accounts were only a subset of the total cases it investigated, but had enough detail to map attacker behaviour against the MITRE ATT&CK framework. I’ll say this is the kind of report that makes security teams read twice, then check their logs again.
AI is helping attackers move beyond basic phishing
Anthropic found that 560 of the 832 accounts, or 67.3 per cent, used AI for malware-related work. That is the noisy part. The more worrying bit sits deeper in the attack chain.
The company said 54 accounts, or 6.5 per cent, used AI to assist with “lateral movement”, which means moving inside a hacked network after getting in. That used to need more skill. Now AI can help weaker actors do harder things.
Anthropic said its analysis had three main conclusions. One of them was clear: “Malicious actors are using AI in ways that make them more dangerous.”
What Anthropic found in banned Claude accounts
| Finding | What it means |
|---|---|
| 832 banned accounts studied | Accounts linked to malicious cyber activity from March 2025 to March 2026 |
| 560 accounts used AI for malware development | Attackers used AI to build or improve harmful tools |
| Medium or higher-risk actors rose from 33% to 56% | More actors moved into higher-risk activity within one year |
| 13,873 malicious actions observed | Activity was mapped across MITRE ATT&CK techniques |
| 482 unique sub-techniques seen | AI misuse appeared across a wide range of cyber tactics |
Why old cyber risk signals may not work now
Security teams usually judge attackers by skill, tools and number of techniques used. Anthropic says that method is getting weaker.
The report says low-skill actors used around 16 techniques on average, while the most skilled used around 20. That gap is not huge. The bigger signal is where AI is used. If the model helps in live operations, credential dumping, account discovery or privilege escalation, the actor becomes more dangerous.
Anthropic also pointed to a state-sponsored cyber espionage operation disrupted in November 2025. In that case, it said the model acted as an autonomous agent that “executed commands, exploited vulnerabilities, stole credentials, and made tactical decisions”.
MITRE ATT&CK may need an AI update
Anthropic said the MITRE ATT&CK framework does not yet fully capture behaviours like autonomous attack chaining, real-time AI decisions and AI-directed execution with little human input.
Anthropic said it is in talks with MITRE on how the framework could evolve.
