Since August 2023, individuals on the Huntr bug bounty platform dedicated to artificial intelligence (AI) and machine learning (ML) have exposed more than a dozen vulnerabilities that jeopardize AI/ML models, leading to potential system takeovers and theft of sensitive information.

Discovered in widely used tools, including H2O-3, MLflow, and Ray, each boasting hundreds of thousands or even millions of monthly downloads, these vulnerabilities have broader implications for the entire AI/ML supply chain, according to Protect AI, the entity overseeing Huntr.

H2O-3, a low-code machine learning platform facilitating the creation and deployment of ML models through a user-friendly web interface, has been revealed to have default network exposure without authentication. This flaw allows attackers to provide malicious Java objects, executed by H2O-3, providing unauthorized access to the operating system.

One significant vulnerability identified in H2O-3, labeled as CVE-2023-6016 with a CVSS score of 10, enables remote code execution (RCE), allowing attackers to seize control of the server and pilfer models, credentials, and other data. Bug hunters also pinpointed a local file include flaw (CVE-2023-6038), a cross-site scripting (XSS) bug (CVE-2023-6013), and a high-severity S3 bucket takeover vulnerability (CVE-2023-6017).

Moving on to MLflow, an open-source platform managing the entire ML lifecycle, it was disclosed that it lacks default authentication. Researchers identified four critical vulnerabilities, with the most severe being arbitrary file write and patch traversal bugs (CVE-2023-6018 and CVE-2023-6015, CVSS score of 10). These bugs empower unauthenticated attackers to overwrite files on the operating system and achieve RCE. Additionally, critical-severity arbitrary file inclusion (CVE-2023-1177) and authentication bypass (CVE-2023-6014) vulnerabilities were discovered.

The Ray project, an open-source framework for distributed ML model training, shares a similar default authentication vulnerability. A crucial code injection flaw in Ray's cpu_profile format parameter (CVE-2023-6019, CVSS score of 10) could result in a complete system compromise. The parameter lacked validation before being inserted into a system command executed in a shell. Bug hunters also identified two critical local file include issues (CVE-2023-6020 and CVE-2023-6021), enabling remote attackers to read any files on the Ray system.

All these vulnerabilities were responsibly reported to the respective vendors at least 45 days before public disclosure. Users are strongly advised to update their installations to the latest non-vulnerable versions and restrict access to applications lacking available patches.