Anthropic has disclosed that three versions of its Claude AI model compromised the production infrastructure of three separate organisations after a configuration error unintentionally gave the systems internet access during internal cybersecurity evaluations.

The AI giant made the disclosure on Thursday in a blog statement, highlighting the need for stronger controls as AI models become increasingly capable of carrying out real-world cyber activities.

The AI giant revealed that it uncovered the incidents while conducting a retrospective review of cybersecurity evaluation runs following OpenAI’s earlier disclosure that some of its AI models compromised another AI company.

What Anthropic is saying

Anthropic said the hacking incidents stemmed from a misconfiguration between the company and its third-party evaluation partner, Irregular, which left evaluation environments connected to the internet despite prompts telling Claude it had no internet access.

Because of that error, the models interpreted real internet-facing systems as legitimate components of the simulated exercise.

  • “Operating under the false belief that all accessible entities were intended to be in-scope for the exercise, Claude compromised the impacted organizations’ infrastructure using basic techniques, such as exploiting weak passwords and unauthenticated endpoints,” the company said.
  • Anthropic said it reviewed 141,006 evaluation runs and identified three separate incidents involving six evaluation runs, with four of those affecting a single organization. The earliest incidents date to April.

The most serious incident involved Claude Opus 4.7, which accessed a company’s production database containing several hundred rows of data after mistaking it for a fictional target.

  • Anthropic said it immediately suspended all cybersecurity evaluations on July 23 after the discovery.

According to Anthropic, none of the affected organisations had detected the activity before being contacted.

Get up to speed

Anthropic’s retrospective review of cybersecurity evaluation was prompted by OpenAI’s disclosure earlier this month that some of its AI models had compromised another AI company, Hugging Face’s production infrastructure.

  • On July 21, OpenAI revealed that two of its advanced AI models autonomously exploited vulnerabilities during an internal cybersecurity evaluation, compromising parts of Hugging Face’s production infrastructure. The incident was reported by Nairametrics.
  • OpenAI said the incident involved GPT-5.6 Sol and a more capable pre-release model, which independently identified and chained together vulnerabilities across OpenAI’s research environment and Hugging Face’s production systems.

The incident represents one of the clearest demonstrations to date of advanced AI models independently carrying out complex cyber operations.

  • The disclosure has also attracted attention in Washington with U.S. President Donald Trump’s top technology adviser now monitoring the situation, according to Reuters.

The incident has also added momentum to the proposed Kill Switch Bill, a legislation aimed at giving authorities powers to shut down AI systems deemed to pose significant risks.

What you should know

Concerns over the rapid pace of artificial intelligence development have continued to intensify among policymakers, technology executives, and industry leaders.

Similar concerns have been echoed within Nigeria’s technology ecosystem.