12,000 Leaked Secrets in AI Training Data Spark Major Security Alarm

A massive trove of nearly 12,000 valid API keys and passwords has been uncovered in the Common Crawl repository—an enormous open-source archive used extensively to train large language models (LLMs). This alarming discovery underscores a systemic risk that could be feeding vulnerabilities directly into AI projects at major tech companies.

Common Crawl’s Global Footprint

Common Crawl has, for over a decade, offered petabytes of freely accessible web data. Because it’s so vast, many AI developers—potentially including those behind models at OpenAI, DeepSeek, Google, Meta, Anthropic, and Stability—rely on these archives to build and refine powerful language models.

Valid Secrets Hiding in Plain Sight

Researchers at Truffle Security analyzed 2.67 billion web pages (about 400 terabytes of data) from the December 2024 Common Crawl archive. Their tool, TruffleHog, flagged a staggering 11,908 secrets that proved fully functional—including AWS root keys, MailChimp API keys, Slack webhooks, and more.

Key Findings

AWS Root Keys: Some were exposed directly in front-end code, granting unrestricted access to cloud resources.
MailChimp Keys: At least 1,500 were hardcoded in HTML or JavaScript, opening the door to phishing campaigns and brand impersonation.
WalkScore API Key: One key alone appeared 57,029 times across 1,871 subdomains, illustrating the scope of credential reuse.
Slack Webhooks: One single webpage revealed 17 unique Slack webhook URLs, which attackers could exploit to post messages or infiltrate private channels.

Why AI Training Data is at Risk

Although LLM training typically involves data-cleaning and filtering to remove sensitive or redundant information, no process is perfect. The sheer volume of data means leaked credentials can slip through, effectively teaching AI systems to regurgitate or even reference them. This threatens the security of the compromised services and raises major concerns about the ethical and legal implications of using publicly scraped data for training.

Consequences of Hardcoded Credentials

By embedding keys into front-end files, developers inadvertently grant anyone scanning the web ready access to powerful backend systems. Attackers could exploit these credentials to:

Spin up malicious cloud infrastructure using stolen AWS resources
Launch phishing campaigns or spam messages via legitimate MailChimp accounts
Infiltrate private Communication channels by abusing Slack webhooks
Exfiltrate sensitive data from compromised servers

Mitigation and Lessons Learned

Truffle Security collaborated with impacted vendors to rotate or revoke thousands of compromised keys. Still, this incident highlights the urgent need for stronger developer practices, especially as AI continues to devour terabytes of public data. Critical steps include:

Never hardcode secrets in front-end files—use environment variables or secure vaults.
Routinely scan codebases with specialized tools (like TruffleHog) to detect exposed credentials.
Implement robust rotation policies so that any discovered key can be rapidly invalidated.
Strengthen AI data preprocessing to better identify and strip out sensitive information before training.

12,000 Leaked Secrets in AI Training Data Spark Major Security Alarm

Related threat feeds

Pakistan Law Enforcement Agencies Targeted in Alleged Multi Group Cyber Espionage Operations

Common Crawl’s Global Footprint

Valid Secrets Hiding in Plain Sight

Why AI Training Data is at Risk

Consequences of Hardcoded Credentials

Mitigation and Lessons Learned

Researcher Chains Three OpenClaw Flaws to Turn a WhatsApp Message into Host Level Code Execution

AssuranceAmerica Breach Leaks 6.99M Driver's Licenses