An Analysis of Misconfigured Sandboxes and Common Vulnerability Classes in the Top Widely Used Agentic AI-enabling Technologies
The enterprise attack surface has been expanded mainly by the proliferation of AI-enabling technologies. Yet, this attack surface is exploited in ways most security teams have not fully reckoned with. Major platforms (e.g., n8n, LangChain, Dify) and AI-powered code editors rely on secure sandboxing to keep potentially harmful actions in check. However, we argue that this reliance on sandboxing warrants a thorough investigation.
In this talk, we will address the following key questions: Q1: How is the sandbox utilized to make the AI-related technology available? Q2: How have the top AI-enabling technologies increased the attack surface and attack vectors? Q3: How have vulnerabilities found in these technologies changed the landscape we've known?
We will present vulnerabilities and insights we've discovered across the top widely used AI-enabling technologies (i.e., Dify, Haystack, LangChain, LlamaIndex, RAGFlow, Flowise, n8n, and VS Code). Although new attack surfaces have been introduced, the underlying root causes remain around long-known classes, such as command injection, path traversal, and insecure deserialization. We will demonstrate common cases that lead to arbitrary code execution when sandboxing is compromised or inappropriately architected, including the following key areas:
How malicious prompts can be treated at the same level as trusted input, which can override the configured system prompt to perform unintended actions Why Python AST-based security checkers are structurally bypassable, and how code execution in a misconfigured deployment can lead to a sandbox escape How unauthenticated internal APIs get exposed on public ports, and why "internal" doesn't mean "protected" in a containerized deployment How server-side template injection in Jinja2 defined templates and dynamic prompts can lead to RCE on the application server
Attendees will leave with an outline of the primary vulnerabilities of AI-enabling technologies, actionable insights for integrating these technologies into systems, and practical strategies to mitigate the corresponding risks.