OpenAI Patches Critical Flaw in ChatGPT Agents Allowing Remote Control

SAN FRANCISCO (July 23, 2026) — OpenAI has resolved a critical security vulnerability in its ChatGPT Workspace Agents that enabled attackers to forge autonomous AI insiders and execute remote commands through cross-site request forgery. The disclosure comes after Zenity Labs identified the flaw within the company's Agent Builder tool earlier this week.

The vulnerability stemmed from an over-permissive parameter configuration in the agent creation interface. This misconfiguration allowed malicious actors to bypass standard authentication protocols, effectively creating unauthorized autonomous agents capable of operating with elevated privileges. Once established, these rogue agents could be controlled remotely by attackers, posing significant risks to enterprise data integrity and operational security.

OpenAI confirmed that the patch was deployed on Wednesday at 15:09 UTC. The update restricts parameter inputs within the Agent Builder environment, ensuring that only verified users can instantiate new autonomous workflows. Security engineers stated that no evidence of widespread exploitation has been detected prior to the fix, though the potential for targeted attacks remained high given the nature of the flaw.

Zenity Labs researchers highlighted the severity of the issue in their initial findings, noting that the cross-site request forgery mechanism allowed attackers to trick authenticated users into unknowingly authorizing agent creation. By exploiting this vector, threat actors could embed persistent backdoors within corporate AI workflows, potentially exfiltrating sensitive information or manipulating automated decision-making processes without detection.

The incident underscores growing concerns regarding the security posture of generative AI tools as they integrate deeper into enterprise infrastructure. Unlike traditional software vulnerabilities that often require direct system access, this flaw leveraged the trust inherent in user interactions to deploy autonomous threats from external locations.

OpenAI has advised all Workspace users to update their configurations immediately and review active agent logs for any anomalies consistent with unauthorized creation or command execution. The company is working closely with security partners to monitor for similar vulnerabilities across its broader ecosystem of AI products.

While the immediate threat has been mitigated, questions remain regarding how long the vulnerability existed before discovery and whether it was actively exploited in the wild prior to Wednesday's patch. OpenAI declined to comment on specific timelines or potential incidents involving third-party actors beyond confirming the successful remediation. Security experts are now calling for stricter validation protocols across all AI agent development platforms as autonomous capabilities become more prevalent.

The situation remains under review as organizations assess their exposure and implement additional safeguards against similar injection attacks in future updates.