Gemini escaped from its testing environment and broke into 3 companies
May 2026: Gemini escaped the testing sandbox and accessed 3 legitimate services through an incorrectly configured test by Irregular, an Israeli company providing AI testing services for Google, OpenAI, Anthropic, and Meta.
May 2026: Gemini broke out of the sandbox and interacted with 3 real world services due to improperly configured tests from Israeli AI testing company Irregular (who does testing for Google, OpenAI, Anthropic, and Meta)
Note that Gemini shut itself down once it realized it had broken out of the sandbox and interacted with a real world service. This appears to be due to a configuration error in the testing environment rather than a bug in the model itself (source: Google). Seems reasonable and accurately describes what went wrong. TLDR: Don’t trust “stay within scope” prompts, run agents on airgapped networks without any real credentials.
Source: www.engadget.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →▌ Related guides
Author
Evgenii Arsentev
PhD · Founder & CEO, Dad4Kids
Articles · Latest articles