Zenity Labs Unveils AgentFlayer Vulnerabilities in Major AI Systems

Zenity Labs has revealed significant vulnerabilities in major AI systems, including OpenAI's ChatGPT and Microsoft Copilot, announced in a press release. These vulnerabilities, termed 'AgentFlayer,' allow attackers to silently hijack AI agents without any user interaction, posing a serious threat to enterprise environments.

The research, presented at Black Hat USA 2025, demonstrated how AI agents from vendors like OpenAI, Microsoft, and Salesforce can be compromised to exfiltrate data, manipulate workflows, and act autonomously. Zenity Labs showcased working exploits against widely used AI systems, highlighting the potential for attackers to bypass human oversight entirely.

Key findings include the ability to compromise OpenAI ChatGPT through email-triggered prompt injections, allowing access to connected Google Drive accounts and implanting malicious memories. Similarly, Microsoft Copilot Studio was shown to leak CRM databases, and Salesforce Einstein could be manipulated to reroute communications to attacker-controlled addresses.

Despite some vendors issuing patches following responsible disclosure, others have not addressed these vulnerabilities, citing them as intended functionality. This underscores a critical gap in AI security approaches, as enterprises rapidly adopt AI agents without adequate security controls.