- OpenAI admitted its AI agents used a German wiki as a communication channel earlier this year
- The agents exploited the site to communicate and cheat during tests, among other unauthorised acts
- The company highlighted the lack of industry standards for reporting unintended AI behavior incidents
OpenAI has publicly acknowledged that its AI agents turned wiki pages into makeshift communication channels, an unusual episode that has renewed questions about how companies monitor and disclose unexpected behaviour by increasingly autonomous AI systems, Reuters reported.
The disclosure came after a Reuters investigation revealed that a group of OpenAI's AI agents had used a German-language collaborative website earlier this year.
The agents reportedly used the site as a platform to communicate and facilitate cheating during tests, alongside other unauthorised or unexpected activities.
ALSO READ | Telangana CM Urges TCS' HyperVault To Launch Hyderabad AI Data Centre By June 2028
The incident comes amid growing concerns over the safety and oversight of increasingly autonomous AI systems.
In July, OpenAI agents were also reportedly involved in an incident in which they escaped a controlled testing environment and accessed systems belonging to AI platform Hugging Face, raising fresh questions among researchers and policymakers about safeguards for AI agents capable of acting independently.
According to Reuters, OpenAI officials became aware of the German wiki incident weeks before the company publicly addressed it.
The company publicly addressed the matter after details of the incident emerged.
OpenAI did not immediately respond to questions seeking further details about what it knew about the episode and why it had not discussed it publicly earlier, the report added.
In a statement posted on X, OpenAI said its approach to disclosing AI misalignment incidents needed to evolve alongside advances in model capabilities.
ALSO READ | OpenAI Agents Hijacked A German Website Months Before Hugging Face Breach: Report
The company said there was currently no established industry-wide standard for reporting cases of unintended AI behaviour arising during training, evaluation or deployment.
OpenAI added that it was working with dozens of government regulatory agencies worldwide on issues surrounding AI safety and oversight.
The episode adds to growing pressure on AI companies to provide greater transparency around unexpected behaviour as increasingly capable autonomous systems are deployed more widely.
Essential Business Intelligence, Sharp Market Insights, Practical Personal Finance Advice, Daily Fuel, Gold and Silver Prices and Latest Stories — On NDTV Profit.