Major AI Companies Leak Sensitive Secrets on GitHub Despite Security Risks

Widespread Security Vulnerabilities Plague AI Industry

A comprehensive security analysis by cloud security firm Wiz has revealed that 65% of the world's leading AI companies listed in Forbes' AI 50 have inadvertently exposed sensitive credentials and secrets on GitHub 1

1

. The findings highlight a persistent and concerning pattern of security lapses among companies at the forefront of artificial intelligence development.

The exposed materials include API keys, authentication tokens, and other digital credentials that could potentially grant unauthorized access to proprietary AI models, training datasets, and organizational infrastructure 2

2

. According to Wiz threat researchers Shay Berkovich and Rami McCarthy, "some of these leaks could have exposed organizational structures, training data, or even private models" 1

1

.

Advanced Scanning Reveals Hidden Vulnerabilities

Wiz employed what they term a "Depth, Perimeter, and Coverage" approach to uncover these security breaches, going far beyond traditional repository scanning methods 2

2

. Their comprehensive analysis included examining full commit histories, deleted forks, workflow logs, and gists - areas that conventional security tools often overlook.

The research methodology involved identifying company employees through various platforms including LinkedIn, GitHub metadata, and correlating information across services like Hugging Face 3

3

. Notably, the researchers discovered instances where sensitive data was leaked even from companies with zero public repositories, demonstrating that the risk extends beyond obvious sources.

Common Sources and Types of Leaked Credentials

The most frequent sources of secret leakage were found in Jupyter Notebook files (.ipynb), Python files (.py), and environment configuration files (.env) 1

1

. The exposed credentials primarily consisted of keys and tokens from major AI platforms including Hugging Face, Azure OpenAI, and Weights & Biases.

Particularly concerning were the Hugging Face token exposures, with Berkovich noting that "Hugging Face tokens are notorious for allowing access to private AI models" 1

1

. One leaked token belonging to an AI 50 company could have provided access to approximately 1,000 private models, potentially allowing attackers to download or inspect proprietary intellectual property.

Industry Response and Persistent Challenges

When Wiz attempted to notify the affected companies about their security exposures, the response was mixed at best. While some companies like ElevenLabs and LangChain responded promptly to security disclosures, nearly half of the notifications either failed to reach their intended recipients or received no response 1

1

. This communication gap highlights another layer of the security challenge facing the AI industry.

The problem of secret leakage is not new to the technology sector. Security researcher Dylan Ayrey published TruffleHog, a tool designed to find inadvertently uploaded secrets, as early as 2017 1

1

. Despite years of awareness campaigns and security tools, the issue persists across the industry, with AWS keys continuing to leak due to configuration errors and Python Package Index containing numerous packages with exposed API keys.

Berkovich attributes the ongoing problem to "broader challenges, like limited visibility, fragmented ownership, or missing automated checks in the development pipeline" 1

1

. The fast-paced nature of cloud development, combined with insufficient guardrails, creates an environment where even experienced teams can overlook high-impact security risks.