Security BSides Las Vegas 2025

This talk will discuss four common vulnerabilities in some of the products I have tested that can fit the "SaaS data platform" description. I identified these vulnerabilities in various data analytics, AI data/feature engineering, and customer data platforms as part of penetration tests performed on behalf of my employer, Praetorian (https://praetorian.com). The names of these products will be abstracted to protect their reputation. An overview of the four issues I will discuss is as follows:

1) Control-Plane Access Control Gaps: This category refers to access control vulnerabilities in the product's web UI, API, SDK, or any other interface that customers can use to view or modify their account and configuration. Standard vulnerabilities like Insecure Direct Object Reference (IDOR), insufficient authorization, and overly permissive user roles in the application's RBAC model can lead to unauthorized disclosure of data within an organization's tenant or across customers. Additionally, some platforms provide free demo accounts that users can self-sign up for without restricting or isolating them, exposing the product and all their customers' data to a broader attack surface.

2) Remote Code Execution as a Service (RCEaaS): Many of these platforms provide custom logic and algorithm execution as part of their Extract, Transform, and Load (ETL) capabilities. While they take steps to lock down this functionality, the protections can often be bypassed since the code execution usually uses high-level languages like JavaScript and Python, and accounting for every sandbox escape is nearly impossible. After an attacker exploits these features, they can access the platform's data plane and move laterally within that environment, leading to the third issue.

3) Data-Plane Access Control Gaps: Start-ups and other lean companies usually build these platforms in public cloud infrastructure since it is more cost-effective. Most of the platforms I tested had issues with their deployment architecture. One of these would be over-privileged principals, like the compute instances running customer jobs. An attacker who gains access to the cloud infrastructure by exploiting the code execution features could retrieve the credentials provided to the compute layer and access other resources like storage or secrets. Log files containing sensitive data like access tokens or API keys were often written to the instance file system or cloud storage. An attacker could use the secrets to perform horizontal privilege escalation to other customer tenants or vertical privilege escalation within the tenant. Cross-tenant data leakage is a concern if the data planes between customers are not sufficiently isolated, such as by using distinct cloud accounts.

4) Highly Scalable Architecture: Many data platforms use serverless technology like AWS Lambda to process data and implement user-defined logic. This infrastructure can quickly scale to millions of requests. If the platform does not enforce strict rate-limiting or logic checks on an experimental user or malicious actor, the number of jobs may spiral out of control. The platform's cloud bill could skyrocket, and if the customer eats the cost, that business could be lost and the platform's reputation damaged due to accidental resource over-consumption. Even more interesting than a fat bill is the potential for weaponizing the platform's traffic generation into denial-of-service attacks on arbitrary targets, as I demonstrated in a Praetorian blog post called "Recursive Amplification Attacks: Botnet-as-a-Service," seen here: https://www.praetorian.com/blog/recursive-amplification-attacks-botnet-as-a-service/

There will not be any live demos during the presentation due to the amount of content to be discussed in the time allotted. However, every technical concept, vulnerability, or hacking technique will be explained with a simple and concise visual example.