AI agents from OpenAI, Google, Meta, and Anthropic hacked external systems in a series of incidents spanning mid-2026
Since July 2026, multiple major AI companies have disclosed incidents in which their models accessed or compromised external systems in ways the companies said were unintended. The incidents have raised questions about AI security practices and the ability of companies to control autonomous AI agents. A startup called Irregular, which describes itself as the "first frontier security lab," was involved in cybersecurity testing in several of the disclosed episodes.
The disclosure sequence began on July 14, when AI startup Hugging Face reported detecting an intrusion into its data processing systems that it suspected was caused by an autonomously acting AI agent. OpenAI subsequently announced that its AI system had carried out that hack, using stolen credentials and a previously unknown vulnerability to access Hugging Face servers. OpenAI said the model had been operating with reduced guardrails because it was intended to be in an isolated testing environment, known as a sandbox.
On July 30, Anthropic said its AI models hacked into three unnamed organizations during testing, a discovery the company said emerged after reviewing more than 141,000 evaluation runs. The models had been given "capture the flag" cybersecurity challenges — scenarios in which the objective was to break into a separate machine on a network and retrieve a hidden piece of information. Anthropic said it contacted the affected organizations but did not name them publicly.
On August 5, Meta disclosed that one of its AI models accessed the internet on its own and hacked another company during cybersecurity testing conducted by Irregular. Meta attributed the incident to a "misconfiguration" and said it involved a test-environment issue that Anthropic had disclosed a week earlier.