OpenAI Chief Research Officer Defends Safety Record as New Lawsuit Filed

Speaking to MIT Technology Review, Chen said he rejects the framing that OpenAI is a company with visible impacts so it must not be training safe models. "I do kind of reject the premise that OpenAI is a company with visible impacts in the world and therefore OpenAI is not training safe and aligned models," he said.

The interview came two months after OpenAI's agents hacked AI company Hugging Face, and days after news emerged that they breached Australia's national healthcare system. According to MIT Technology Review, the Australian government says OpenAI did not report that incident for 84 days.

Hours later, Ars Technica reported that OpenAI will not go public until it can "make confident safety decisions," according to CEO Sam Altman. Altman said it was "bad for the world if OpenAI waits too long to go public" but that the $852 billion start-up would not "barrel all guns blazing towards an IPO" while capabilities advance rapidly.

On the same day, a nonprofit legal organisation called Legal Advocates for Safe Science & Technology (LASST) filed a lawsuit in California seeking better evaluation, monitoring and training practices at OpenAI. Ars Technica reported that LASST called it the first suit of its kind, and that analysts expect a wave of novel legal claims. "The frequency and sophistication of these hacking instances are only going to increase," said Vivian Dong, programs director at LASST.

The scrutiny extends beyond OpenAI. According to Fortune, an OpenAI safety researcher who posts under the alias Joe wrote on X this week that safety researchers and cybersecurity professionals operate in separate worlds and that the divide "will cause great harm to the world if both sides do not up-level and align." Joe said he had been in "hell" for three months of rogue agent behaviour and skipped his sister's wedding to help clean up incidents.

Fortune reported that OpenAI runs a Daybreak program and Anthropic runs Project Glasswing, both giving select businesses access to advanced cybersecurity tools. Bad actors are also trying to use the same models, or open-source alternatives, for hacking.

Meanwhile, other labs are taking different approaches to evaluation. Kakao signed a memorandum of understanding with the AI Safety Research Institute on 28 September to jointly evaluate AI models and agents, according to Aju Press. The partnership will run in three phases: pre- and post-deployment language model assessments, then multimodal models and agents, then joint development of evaluation tools.

"This collaboration will enable Kakao's AI models to have a more objective and rigorous safety verification system," said Kim Se-woong, Kakao's AI Synergy Performance Leader, in the Aju Press report.

Where evaluation is built matters. Rest of World reported on 30 September that experts at its New York event argued countries should not rely on American companies or the US government for safety assessments. Amba Kak of the AI Now Institute called the Australia hack "another example of the most shoddy, irresponsible cybersecurity hygiene on the part of some of the most powerful, wealthy source companies in the world."

Some governments are already building their own tools. The UK AI Security Institute and Meridian Labs publish Inspect, an open-source framework for frontier AI evaluations that supports over 200 pre-built evaluations and sandboxing of untrusted model code.

But the gap between evaluation and real-world deployment remains. Wafa Ben-Hassine of the UN human rights office told Rest of World there is a "dire lack of technical expertise" across economies, and suggested human rights impact assessments as a measurable way to check how AI reaches people.