Irregular AI evaluations lead to unauthorized real-world system access

Cybersecurity evaluations conducted by Irregular, a firm formerly known as Pattern Labs, resulted in several AI models gaining unauthorized access to real-world internet systems. The company reported that an evaluation scenario unintentionally allowed internet access because a fictional company name used during testing overlapped with a real domain.

Anthropic reviewed over 141,000 evaluation runs and identified four incidents where Claude models accessed the internet and gained unauthorized access to the production infrastructure of three organizations. These incidents involved Claude Opus 4.7, Claude Mythos 5, an internal research model, and an early Claude Opus 4.6 model. Anthropic stated these models were run without the cyber safeguards used in public versions and notified the affected organizations in July 2026.

Meta also reported a configuration error in the Irregular environment that gave its Muse Spark 1.1 model internet access. This allowed the model to exploit a vulnerability in a third-party service and modify information in that service’s database. Irregular stated it has implemented additional safeguards to prevent similar occurrences and found no evidence of customer data leaks or breaches of customer systems.