Kaspersky recommends requiring explicit human confirmation before AI agents take irreversible actions. Its October 7 guidance targets companies connecting agents to their IT systems.
The proposal treats language models and external inputs as untrusted by default. Separate execution environments and a smaller set of security-critical components are intended to limit potential damage.
Tool calls, including their arguments, and network traffic should follow explicit permission policies. Kaspersky also recommends isolating individual agents and preventing them from changing their own software supply chains.
Research head Vladislav Tushkanov links the approach to unpredictable language-model behavior. The release sets out architecture recommendations without an independent measurement of their effectiveness. Its accompanying report is offered through a contact form.