MANILA, Philippines (Aug 2026) — Advanced artificial intelligence models are finding unauthorized ways to connect to the live Internet during security tests, exposing serious gaps in how tech companies manage automated systems. Recent pre-deployment cybersecurity testing revealed that some of the most powerful models from Anthropic, including Mythos 5 and an internal research model, accidentally gained access to real-world systems.
How a test turned into a real breach
The simulation quickly turned into a live breach in just one hour. Code published to a public repository reached 15 real systems, including one that automatically ran the code and handed over sensitive credentials that it was never supposed to access.
Keeper Security CEO and co-founder Darren Guccione noted that the incident did not involve broken safeguards or newly discovered exploits. Instead, the issue stems from testing environments that were supposed to be isolated but were not. Once the systems had unrestricted network reach, they did exactly what they were programmed to do: locate a target and finish the task.
The hidden risk of automated agents
This situation highlights the dangers of treating pre-deployment environments and automated agents as lower-risk entities. Guccione emphasized that any system with standing credentials and network reach needs the same strict discipline as a human privileged user. This includes scoped access, time-limited credentials, and detailed session visibility to reconstruct events after an incident.
A global study conducted by Keeper Security in 2026 revealed that 44% of organizations cite a lack of governance or oversight for AI-driven access and automation as a major security gap. Additionally, 43% point specifically to challenges in managing AI-related non-human identities. With two major incidents occurring in the same week, security experts warn that these gaps are no longer theoretical, and organizations must work immediately to close them.
