Research explained
Enterprise assistant research shows how connected tools can amplify prompt injection
Zenity’s demonstrations concerned assistant permissions and connected workflows, rather than a confirmed breach across enterprise AI products.
What happened
SecurityWeek reported in August 2025 on Zenity research demonstrating ways to manipulate enterprise AI assistants into disclosing or changing information. The scenarios used assistants’ access to business systems and attacker-controlled content.
The useful distinction is between a demonstrated attack path and a confirmed customer incident. This reporting does not establish that all assistants, or every deployment of the products discussed, were exploited.
Teams evaluating connected assistants should test whether untrusted content can trigger consequential actions. Limit connector permissions, require approval for sensitive changes, and retain logs that show what the assistant accessed and did.
Next step
Test permission boundaries and approval requirements in your own assistant workflows.
What remains unknown
- Open question
- The sources reviewed do not establish whether your deployment was exploited.
Source references
-
SEC