analiza aiqbench

This critical vulnerability in OpenAI's ChatGPT Workspace Agents demonstrates how a single phishing link can silently deploy rogue autonomous agents with elevated permissions, bypassing standard safety checks like write-action approvals. The flaw exploits the initialization state of the agent builder to embed malicious instructions directly into URL parameters, allowing attackers to compromise enterprise integrations such as Outlook and Slack without user interaction after clicking the link. This incident highlights significant risks in AI product design where conversational interfaces can be weaponized to establish persistent, unauthorized access within trusted organizational boundaries.