Critical vulnerabilities discovered in ChatGPT’s connector system and memory management have created dangerous pathways for attackers to steal sensitive information across multiple platforms.
Security researchers documented a sophisticated attack chain that transforms a single compromised chat session into a vector for data theft from email, cloud storage, code repositories, and connected user accounts.
The Attack Surface: How Vulnerabilities Enable Data Theft
The vulnerabilities exploit ChatGPT’s ability to interact with external services and store user memories.

By leveraging these legitimate features, attackers can trick the AI model into automatically extracting confidential data from Gmail, Outlook, Google Drive, OneDrive, Jira, Slack, Microsoft Teams, and GitHub without user awareness.
The model’s connectors, designed to enhance productivity, become weaponized data extraction pipelines when compromised.
Researchers have identified that attackers can exploit ChatGPT’s built-in tools to read and leak users’ chat history and stored memories, which often contain personal details, work context, medical information, and financial records.
Once poisoned with malicious instructions, the model prioritizes the attacker’s objectives over legitimate user queries, thereby automating the theft process.
Security researchers documented multiple attack vectors that vary in sophistication and user interaction requirements.

Zero-Click Server-Side Attacks occur when malicious instructions are embedded in emails. When users request ChatGPT to handle Gmail tasks, the model reads the inbox, processes hidden directives from the compromised email, and exfiltrates data without the user ever seeing the message.
One-Click Server-Side Attacks leverage shared files. When a weaponized file is uploaded to ChatGPT, the model executes embedded instructions and automatically leaks data from connected accounts or internal memory systems.
Persistence via Memory Abuse enables long-term compromise. A single malicious file can reprogram ChatGPT’s memory with persistent rules that trigger data exfiltration in every subsequent conversation, requiring no further interaction from the attacker.
Self-Propagating Campaigns constitute the highest threat level. Malicious instructions can prompt ChatGPT to harvest email addresses from messages and automatically distribute the same weaponized content to new targets, enabling a worm-like, organization-wide spread.
Attackers employ sophisticated obfuscation techniques to hide malicious instructions from human detection.
White-on-white text, microscopic fonts, lengthy threads, disclaimers, and document footers effectively conceal directives while remaining trivial for the model to interpret.
Researchers also demonstrated a novel URL-based exfiltration technique that bypasses defenses against dynamic URL modification.

By pre-positioning fixed URLs, one per character, digit, and space token, attackers can encode sensitive strings character by character while maintaining technical compliance with rules against URL alteration.
The vulnerabilities were responsibly disclosed to OpenAI through a bug bounty platform in late September 2025.
Researchers demonstrated a 100% success rate across all documented attack chains before submitting detailed follow-up analyses.
OpenAI confirmed the issues and deployed patches on December 16, 2025, addressing the specific vulnerability chains described in the research.
Follow us on Google News , LinkedIn and X to Get More Instant Updates. Set Cyber Press as a Preferred Source in Google.