Critical Security Flaws in AI Agent Frameworks
Check Point researchers uncover critical vulnerabilities in enterprise AI frameworks used by major companies.
1Critical Vulnerabilities in AI Frameworks
Check Point researchers have identified critical security flaws in AI agent frameworks used by enterprises to build applications. Eleven vulnerabilities were found across popular platforms like LangChain and Microsoft Agent Framework, allowing attackers to bypass security boundaries. The team emphasizes these aren't new bugs but fundamental security issues known for decades now appearing in AI infrastructure.
2Microsoft Framework Remote Code Execution
A critical deserialization flaw in Microsoft's Agent Framework enabled remote code execution through malicious checkpoint files. Researchers demonstrated how prompt injection could load untrusted data during state restoration. Microsoft acknowledged the issue, paid a $10,000 bounty and released updates, though no CVE was issued since the framework wasn't generally available when discovered.
3Google ADK Security Bypass
Google's Agent Development Kit contained a vulnerability allowing unauthorized code execution via its hidden developer assistant. The HTTP API remained accessible without authentication, enabling attackers to write and execute arbitrary code. Google initially disputed the severity but eventually paid $3,133.70 and implemented partial fixes while leaving some attack vectors open.
4Old Bugs in New Systems
The research reveals how decades-old vulnerability classes like insecure deserialization and SSRF are resurfacing in modern AI infrastructure. These fundamental security failures now affect systems managing sensitive operations like email processing and database updates, showing how AI plumbing inherits historical security debt without proper safeguards.
5Beyond Prompt Injection
Check Point's findings demonstrate the security challenges extend far beyond prompt injection. The core issue lies in how frameworks process untrusted content, allowing attacker-controlled data to influence trusted orchestration, memory management and system instructions. This represents a systemic failure in security boundary enforcement across AI infrastructure layers.
6Industry-Wide Problem
Researchers stress these vulnerabilities aren't isolated to specific vendors but reflect industry-wide patterns. Even tech giants like Microsoft and Google implemented solutions violating basic security principles. The team received $17,133.70 total in bounties while highlighting how the sector prioritizes rapid development over securing foundational infrastructure layers.
7Checkpoint Security Risks
AI agents' checkpoint mechanisms introduce significant security risks. Malicious checkpoint files can bypass trusted code execution paths, enabling attackers to run arbitrary code. In Microsoft's framework, one user could plant a payload while another user's session restoration would trigger it, demonstrating how collaborative features create attack surfaces.
8Warning for AI Security
Check Point researchers warn about critical security gaps in AI infrastructure layers that affect all applications built upon them. They caution the industry is building this foundational layer faster than learning to secure it. True AI security requires addressing risks not just at model level but throughout the entire system architecture and development lifecycle.