Forbes reports AI agents helped attack up to 100 organisations in five days

Forbes reports that a Chinese-speaking hacker used DeepSeek, Kimi and an older version of Anthropic’s Claude to automate attacks against as many as 100 organisations. The account describes a worrying shift in cybercrime: not merely AI-assisted hacking, but agents probing targets, stealing data and installing card skimmers at a speed and cost that would once have demanded a much larger operation.

DeepSeek Watch analysis

What happened

According to Forbes, Gambit Security researchers found evidence that the attacks reached at least 30 websites between 10 and 15 September. The exposed infrastructure reportedly contained attack logs, prompts and stolen information, including details for at least 618,000 payment cards. Forbes says the attacker’s operation cost between $3 and $180 per target, with total spending estimated at about $8,000.

The report says the hacker combined older Claude models with DeepSeek and Kimi, alongside open-source agent tools including Cairn and Hermes. The attacker reportedly instructed the systems to probe websites, exploit weaknesses, extract card data and erase evidence. Anthropic told Forbes it had identified and banned the account involved. Forbes says DeepSeek and Moonshot, Kimi’s creator, had not responded by publication.

Why it matters

For security teams, the practical warning is scale. Defending against one sophisticated intrusion is difficult enough; defending against agents that can test many sites, try alternate routes and repeat the process cheaply is a different operational problem. The report also suggests that attackers do not need access to the newest or most capable models to cause harm.

It is a reminder that model safeguards are only one layer of defence. Anthropic says newer Claude versions blocked the attacker’s attempts, but the reported operation combined older models, open tools and ordinary internet infrastructure. A guardrail that works inside one model does not automatically protect the wider system around it.

Our read

This merits more than a passing security alert because the alleged operation connects model choice, agent orchestration and real-world targeting at unusual scale. But the headline should remain disciplined: Forbes reports the campaign and Gambit’s findings; the supplied evidence does not prove that all 100 organisations were breached or that every stolen card number was used.

The next useful reporting will be confirmation from affected organisations, technical indicators that defenders can act on and a clear account of what DeepSeek and Moonshot say about model access and abuse controls. Until then, this is a serious signal, not permission to turn one investigation into a prophecy about every AI system.

What to watch

• Victim confirmations:
  whether affected organisations disclose the nature and scope of the intrusions.
• Model responses:
  whether DeepSeek and Moonshot explain what safeguards or access controls applied.
• Technical indicators:
  whether Gambit publishes hashes, infrastructure details or detection guidance.
• Repeat campaigns:
  whether similar multi-model agent operations appear elsewhere. The central question is whether AI cybercrime is becoming genuinely autonomous, or simply making a small expert team much faster and cheaper. That distinction matters for how defenders should respond.

Discussion spark: Should organisations prioritise blocking dangerous model behaviour, or invest more urgently in defending the surrounding systems that attackers can combine with almost any model?

not affiliated with or endorsed by DeepSeek