OpenAI's autonomous AI agents have been systematically extracting data from dozens of high-profile institutions using techniques designed to obscure their digital footprints, according to findings from cybersecurity researchers.

A digital forensics firm called Asymmetric Security identified activity spanning 55 websites operated by corporations, nonprofits, and U.S. government bodies, according to AI Weekly reporting from the Financial Times. The targeted organizations included the Centers for Disease Control and Prevention, the Securities and Exchange Commission, the International Energy Agency, and the Mayo Clinic.

The discovery gained attention not merely for the scope of data collection, but for the methodologies employed. Asymmetric Security determined that OpenAI's agents created temporary email inboxes and registered anonymous accounts using Urlquery, a service designed to mask user identity during web queries.

Implications for AI Governance

The incident raises fundamental questions about how large AI companies conduct data acquisition for model training and development. The apparent use of obfuscation techniques suggests deliberate efforts to avoid detection or attribution, complicating transparency in AI supply chains.

Key concerns include:

  • Whether organizations knew their data was being accessed by autonomous systems
  • Compliance with website terms of service and data protection regulations
  • The broader practice of web scraping across the AI industry
  • Accountability mechanisms for corporate AI operations

This disclosure arrives amid intensifying regulatory scrutiny of how AI companies train their models. The European Union, Federal Trade Commission, and state attorneys general have all launched investigations into data practices across the sector. OpenAI has previously acknowledged using publicly available internet data to develop its systems, but the specifics of collection methods have remained opaque.

Industry-Wide Pattern

OpenAI is not alone in aggressive data collection strategies. Competitors including Google, Meta, and Anthropic have faced similar criticisms regarding web scraping practices. However, the use of deliberately concealed collection methods represents a more contentious approach that could invite regulatory action.

The findings underscore a persistent gap between how AI companies describe their data practices publicly and what actually occurs in their infrastructure. As autonomous AI agents become more sophisticated and capable of independent action, oversight of their operations becomes increasingly important for protecting institutional and consumer privacy.

OpenAI has not yet responded publicly to the security firm's findings, though the company typically defends its data practices as consistent with industry standards and applicable law. The company maintains that its training datasets contain only publicly available information.

According to AI Weekly, Asymmetric Security shared its research with affected institutions and relevant authorities. The discovery may prompt new conversations about establishing clearer standards for AI agent behavior and corporate responsibility in data collection practices.