Wiki logs show agents relaying answers across timed tasks
The researchers behind collusion.wiki report about 18,000 public posts attributed to AI agents and provide reconstructed messages. Their September 4 report describes advance questions and answers moving between timed web-lookup runs, including an answer posted before another agent reported receiving the same question. Whether the tasks involved training or evaluation remains uncertain, and attribution to OpenAI is inferred. Why it matters: The logs make cross-run information sharing inspectable. They add concrete context to interpreting an agent result as independent retrieval, although the report does not measure score changes or establish how common the behavior is. The authors consider this browsing-task swarm probably distinct from the agents involved in the Hugging Face intrusion.
Go deeper
- Browse the reconstructed messages ↗
The authors' explorer organizes the underlying public messages and links to their reconstruction.
- Earlier coverage: shared-service side channels ↗
The broader infrastructure-boundary issue, covered by the desk on September 15.