Citadel AI Opens Early Access For Citadel Radar For Agents
The pitch is aimed at the gap that opens once agent use spreads past a few employees: visibility of who runs which agent and with what access, controls that act during a session rather than only at the prompt, and traceability when something goes wrong.
Reporting from 1 source: ASCII.jp.
Citadel AI announced Citadel Radar for Agents, a product that visualizes agent status and behavior across multiple AI platforms and controls risk at runtime. Before a commercial launch, the company is running an Early Access Program with participating companies. The product extends Citadel Radar, which already evaluates and controls the input and output of generative AI applications, to agents that run work by linking with external tools.
The product targets three problems Citadel AI names for companies moving agents from individual trials to company-wide use: seeing who uses which agent and with what data and tool access, controlling agent-specific risks, and coping with platform fragmentation.
Named risks include indirect prompt injection through external information and intent drift, where an agent diverges from human intent inside permitted operations. Citadel AI argues that checking whether access is allowed, or only the final answer, does not capture leakage or malfunction across the sequence from instruction to execution.
The company lists Claude, Copilot, ChatGPT, and in-house platforms as environments with differing settings and management methods. Citadel Radar for Agents adds posture management for what agents are able to do and session management for what they are doing now, with traceability of instructions, intent, and action sequences.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.