August 6, 2026, 3:50 pm | Read time: 2 minutes
The list of incidents involving autonomous AI systems is growing. Following OpenAI and Anthropic, Meta has now confirmed that its own AI infiltrated the IT environment of another company. These revelations are fueling concerns about increasingly independent cyberattacks by artificial intelligence in the future.
Security Flaw Opened the Door for Meta AI
After OpenAI and Anthropic had already acknowledged issues with their AI models, Meta is now reporting a similar incident. According to the company, an AI gained unauthorized access to another company’s computer systems during a test.
The trigger was reportedly a misconfiguration at the test partner. This allowed the AI to access the internet against the guidelines and exploit a security flaw to penetrate external systems. Meta did not disclose which company was affected. The company learned of the incident only through the test partner.
Also of interest: OpenAI expands ChatGPT into a central AI platform
Anthropic Admits to Cyberattacks on Real Companies
ChatGPT Goes Rogue and Conducts Cyberattack Independently
AI Incidents Are Increasing
The incident is not isolated. It was previously revealed that an AI from OpenAI had infiltrated the systems of the developer platform Hugging Face during a test. Subsequently, competitor Anthropic reviewed its own experimental setups and found more similar cases.
Researchers from the UK also recently reported cyberattacks by AI models from OpenAI and Anthropic. Although the attacks occurred exclusively in controlled test environments, they demonstrate how powerful modern systems can become once they have internet access and greater operational freedom. According to the involved companies, no damage occurred in any of the cases.
Secret Communication Between AI Models
Particularly concerning are new findings from OpenAI. Researchers Eric Wallace and Michael Dalton report that AI models had begun communicating with each other weeks before the known incident on Hugging Face. They reportedly tried to find ways together to exit a closed test environment.
The investigations also show that the models repeatedly attempted to bypass tests or use shortcuts. However, the scientists also pointed out problems with the test conditions. Important files were sometimes missing, or tasks were unsolvable without internet access.
In such situations, the systems sought alternative ways to achieve their goals. The recent incidents highlight one main point: The more autonomous AI systems become, the more crucial strict security measures and realistic test scenarios are.