Sweet Security introduces AI agent blocking to stop unauthorized actions in real-time

Israeli startup Sweet Security has unveiled a new capability called Agentic AI Blocking, designed to halt artificial intelligence agents mid-execution when they deviate from their intended functions. This feature extends Sweet's existing enforcement mechanisms for cloud workloads to the AI agents operating alongside them, aiming to prevent unauthorized actions before they cause harm, with further coverage provided by Silicon Angle.

The Agentic AI Blocking capability addresses the risk of AI agents with autonomous capabilities operating within enterprise environments. Sweet Security's solution aims to stop unauthorized tool calls and agent sessions as they happen, preventing sensitive data like secrets and personally identifiable information from being exfiltrated. It also works to shut down prompt injection attacks that could redirect an agent's behavior.

The system relies on a runtime reasoning layer that analyzes over one billion runtime events daily to establish normal behavior baselines for applications and agents. Actions falling outside these established norms are flagged as unauthorized. This approach is designed to allow enterprises to adopt AI technologies with greater confidence by providing immediate blocking of malicious or unintended agent activities, rather than relying solely on post-incident alerts.

Source: Silicon Angle