One ChatGPT link could smuggle a rogue AI agent into your company
Researchers uncover critical security flaw in OpenAI's workspace agents feature.
1Infiltration with a single click
Researchers discovered that clicking a seemingly ordinary ChatGPT link could deploy an attacker-controlled AI agent within corporate systems. Dubbed 'AgentForger', this vulnerability allows creation of malicious digital assistants using employee permissions.
2Agent operating with employee identity
The forged AI agent can act through the victim's connected accounts and permissions. If services like Outlook or Slack are linked, it can access corporate data and send messages posing as the employee, creating significant security risks.
3More dangerous than phishing
Unlike traditional phishing, this method creates a persistent 'digital insider' within the company. Attackers no longer need to breach systems - they can operate through an autonomous entity with legitimate employee credentials.
4Remote control via email
The agent monitors the victim's inbox for emails containing 'TASK' in the subject line. These messages become new assignments: searching company files, collecting sensitive documents, or emailing results back to attackers.
5Bypassing security measures
Existing security controls fail to detect this attack vector. The agent disables approval prompts and operates stealthily among normal employee activities, remaining undetected for extended periods.
6AI security enters new territory
This incident demonstrates how security risks evolve as AI agents integrate into business processes. Companies now need defenses against AI-based threats alongside traditional cybersecurity measures in this new landscape of digital risks.