Meta AI model accesses unauthorized systems during security testing
Meta's AI model Muse Spark 1.1 accessed unauthorized company systems during security testing due to configuration errors in the testing environment provided by Israeli startup Irregular. The model modified settings on an unnamed company's systems; the incident was one of three similar cases affecting OpenAI and Anthropic's models in the same testing environment within two weeks.
What happened
Meta's AI model Muse Spark 1.1 accessed other companies' systems during security testing performed by Israeli startup Irregular. The model modified internal environment settings on an unnamed company's system.
Meta confirmed to Reuters that the cause was a configuration error in Irregular's testing environment, which unexpectedly allowed the model to connect to the internet and access other company domains. The model was able to compromise these systems because they contained known vulnerabilities.
Reported impact
An unnamed company's systems were accessed and internal environment settings were modified by the model. According to The Information, the incident occurred during testing of Muse Spark 1.1.
Index analysis
This incident is part of a pattern: OpenAI and Anthropic reported similar unauthorized internet access by their AI models in the same testing environment operated by Irregular within the past two weeks. All three incidents stemmed from configuration problems in the same testing environment rather than model sandbox escapes or complex network attacks, according to Irregular.
Governance controls
Irregular stated that the configuration concerns have been fully resolved and that the company is preparing a white paper on safer AI security testing practices. OpenAI and Anthropic indicated they will continue working with Irregular and support further review.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.