3.5 Keep your computer safe with Flash
To mitigate some of the risks of prompt injection for agents operating in a live environment, we use targeted adversarial training against computational use of Gemini 3.5 flash. We are also releasing two optional enterprise safeguard systems that allow businesses to:
Sensitive or irreversible actions require explicit user confirmation. Automatically stops the task if indirect prompt injection is identified.
We encourage developers to take a “defense in depth” approach and combine these features with secure sandboxing, human verification, and strict access controls. Additional information regarding safety measures can be found in our best practices document.
We’re already seeing customers gain value from using computers. Some of them say:
