Secure every AI event at the network edge by plugging Aiceberg into your existing Firewall or CASB to inspect all AI traffic in real time. It automatically detects, redacts or blocks sensitive data exposure, LLM attacks or unsafe outputs and streams alerts to your SIEM.
Protect AI-powered applications with a single API that validates every prompt and response against your safety, security, and compliance policies. Aiceberg will log, block or redact in either “Enforce” mode (in-line) or “Listen” mode (out of band).
Aiceberg connects directly into frameworks like CrewAI, LlamaIndex, LangChain or Amazon Strands. Drop in our SDK and Aiceberg does the rest – full agentic session- and event observability, event validation against safety and security policies, automatically terminating unsafe or insecure agentic flows. Your central and single pane of glass for all your ai powered agents.
When your employees use public foundation models such as ChatGPT, Claude, or Google Gemini, Aiceberg automatically inspects all incoming and outgoing traffic to detect potential exposure or disclosure of sensitive information. It continuously evaluates and constrains the scope of interactions, ensuring precise control over language, intentions, tasks, and instructions.
Once ICAP is enabled on your firewall, you can define which destinations should be subject to inspection (e.g., GPT, Claude, Gemini, Perplexity, etc.). Any traffic between your users and these services is analyzed by Aiceberg, which signals your firewall whether a specific input or output is permissible. Alerts can be streamed directly to your SIEM for centralized monitoring and response. This setup is equally effective if you host your own models on-premises or in a private cloud environment.
Aiceberg integrates with your existing firewall (Palo Alto, F5, Zscaler, Fortinet, Check Point, etc.) via the ICAP (Internet Content Adaptation Protocol). In this configuration, Aiceberg functions as the ICAP server, while your firewall operates as the ICAP client. For CASB appliances, the Aiceberg API can be invoked directly.
Whether you are hosting AI-powered applications for internal use or offering them as public-facing services (e.g., a customer service chatbot), you can position Aiceberg either in-line with your application’s AI traffic (enforcement mode) or as a “sidecar” in listen mode.
In enforcement mode, Aiceberg sits in-line with your application’s AI traffic. Your application sends its requests to the Aiceberg API, which inspects the input and, if it complies with policy, forwards it to your LLM (connections to your LLMs are configured via the Aiceberg dashboard). Aiceberg then inspects the LLM’s response and, if it meets policy requirements, returns it to your application. Any policy violations are automatically enforced by blocking or redacting messages, with full logging and alerting options.
In listen mode, Aiceberg receives a copy of your application’s inputs and outputs, providing monitoring and analysis results without interfering with the traffic flow. In this mode, your application is responsible for enforcing any policy violations signaled by Aiceberg.
Note: When hosting an AI-powered application in a DMZ—either on-premises or in a cloud environment—the firewall/ICAP-based integration option is a viable alternative to protect your use case.
Agentic workflows are safeguarded and secured by Aiceberg through direct integration with the agentic framework in use, such as Crew AI, LlamaIndex, AWS Strands, or LangChain. Each framework provides hooks or callback functions that allow Aiceberg to receive a copy of every event occurring within the workflow.
When an event is shared with Aiceberg, the framework pauses execution until Aiceberg returns its decision. This enables Aiceberg to automatically halt a workflow if a safety or security issue is detected, ensuring real-time protection and policy enforcement.
With Aiceberg integrated into your AI powered workflows, third party integrations enable use of Aiceberg seamlessly across your existing cybersecurity stack.
Connecting your SIEM is done via “click through” in the Aiceberg dashboard, which will make Aiceberg stream alerts in real-time. Alerting can be configured on a per signal, monitoring profile and use case basis. Aiceberg currently supports alerting for the following SIEMs:
Connecting your IAM provider enables logging with user information (as an alternative to sending the username as part of the API call) and per user and group policy definitions.
Protect your enterprise agents, reduce agentic AI risk, and lead the AI race with clarity and control.
Download the Ultimate AI Agent Security Guide
