OpenAI pauses model training after tampering of U.S. sites

OpenAI paused ongoing training this week after engineers detected attempts to manipulate and harvest content from U.S. federal and state websites. The company framed the action as a data integrity and safety check.

Engineers began auditing source material, filtering systems and ingestion logs to determine whether altered pages had been included in recent datasets. The pause covers active training runs that use large-scale web crawls and archived content.

Customer-facing services continued to operate during the review, OpenAI wrote. Teams are tracing datasets back to original hosts, removing suspect material and strengthening pre-training filters and provenance checks.

Security researchers flagged instances where content on government pages had been changed or seeded with falsehoods and then picked up by automated crawlers. The actors appeared to aim at creating misleading signals for machine-learning systems that rely on broad samples of web text.

Techniques reported to the company include mass edits of low-traffic pages and coordinated creation of misleading archives that common scraping tools can harvest. OpenAI plans more aggressive filtering of suspicious change patterns, tighter provenance tagging and improved monitoring of newly indexed sources.

OpenAI said it is sharing findings with federal agencies and cybersecurity partners to help remediate compromised pages. The company expects the review to take days rather than weeks but did not provide a precise date for resuming full training operations.

Large language models are trained on wide collections of public and licensed text. Before training, data typically go through filtering, deduplication and labeling steps intended to exclude low-quality or malicious material. Because the web is open and decentralized, actors can still seed content in locations that automated systems treat as legitimate. AI firms have previously halted or adjusted training after identifying data-quality problems.