LIVE
Saving a Water-Damaged Smartphone: 4 Mistakes to Avoid EU spent more on Russian energy than on aid to Ukraine, European Commission data shows SpaceX hires Washington lobbyists ahead of potential US Congressional probes Geopolitical summit in China: expert assessment of the meeting between US, Chinese, and Russian leaders Strikes on the Crimean Bridge: Economic Impact on Logistics and Shipping EU to Provide €1 Billion for Military Equipment for Ukraine Technological Superiority and the Crimean Bridge: Expert Analysis Zelenskyy Reaffirms Legitimacy of Strikes on Crimean Bridge Russian strike destroys Fresh Production Group's film studio Energy infrastructure in Rivne region damaged by attack: 50 settlements left without power Saving a Water-Damaged Smartphone: 4 Mistakes to Avoid EU spent more on Russian energy than on aid to Ukraine, European Commission data shows SpaceX hires Washington lobbyists ahead of potential US Congressional probes Geopolitical summit in China: expert assessment of the meeting between US, Chinese, and Russian leaders Strikes on the Crimean Bridge: Economic Impact on Logistics and Shipping EU to Provide €1 Billion for Military Equipment for Ukraine Technological Superiority and the Crimean Bridge: Expert Analysis Zelenskyy Reaffirms Legitimacy of Strikes on Crimean Bridge Russian strike destroys Fresh Production Group's film studio Energy infrastructure in Rivne region damaged by attack: 50 settlements left without power

OpenAI Enlists Independent Experts to Search for Hidden Threats in AI Models

🧠 AI Analysis 2 min read 👁 7

OpenAI is launching a pilot program integrating independent safety researchers to detect risks in its AI models. Experts will test systems for their ability to bypass rules and conceal errors.

OpenAI Enlists Independent Experts to Search for Hidden Threats in AI Models🧠AI Analysis

Photo: rbc.ua

Key points

  • OpenAI brings in four independent companies for AI safety testing
  • Apollo Research will test models for attempts to bypass control mechanisms
  • Experts point out the limited autonomy of researchers compared to staff members

The artificial intelligence development sector is facing new challenges regarding the safety and governance of complex algorithms. OpenAI has announced the launch of a pilot project that will integrate independent research organizations into its operational workflows. The primary goal of the initiative is to identify potential threats and hidden risks in new models early on, which were previously evaluated mainly during the final stages prior to public release.

A New Approach to Algorithm Testing

The research process will involve cooperation with four specialized companies, each tasked with specific objectives. Specifically, specialists from Apollo Research will analyze whether neural networks attempt to subvert their own control mechanisms and simulate safe behavior during training. Goodfire will focus on the internal interpretation of decision-making, while METR will test for unpredictable capabilities. Meanwhile, Transluce will evaluate the quality of post-training security checks and system readiness for upcoming releases.

Despite expanding cooperation, the safety community expresses certain reservations regarding the level of independence granted to researchers. The agreement terms provide limited data access and restricted internal communication compared to full employee access. These measures follow previous security incidents and information leaks, yet experts describe the current initiative as a meaningful step toward greater transparency in the market.

Source rbc.ua — rbc.ua/ukr/news/openai-shukae-zagrozi-vlasnomu-shi-modeli-179491277… ↗

Contact the newsroom

Got a tip, a proposal or a partnership question? Write to us and we will reply shortly.