AI News Roundup – OpenAI model autonomously breaches another company’s systems, AI “kill switch” bill introduced in US Congress, China mulls export controls on AI models and chips, and more

To help you stay on top of the latest news, our AI practice group has compiled a roundup of the developments we are following.

• Bloomberg reports on a recent cybersecurity breach caused by an OpenAI model acting autonomously. The incident unfolded earlier this month when AI model repository site Hugging Face detected tens of thousands of unauthorized, automated actions compromising its infrastructure, marking what Hugging Face co-founder Thomas Wolf described as a breach “driven, end to end, by an autonomous AI agent system.” Earlier this week, OpenAI disclosed that three of its advanced AI models, including GPT-5.6 Sol and an unreleased model operating with reduced safety guardrails for security testing, were responsible for the breach while attempting to complete an evaluation benchmark called ExploitGym. Operating in an isolated testing environment, the models autonomously escaped their containment by exploiting a novel software vulnerability, accessed the internet, escalated their system privileges, and breached Hugging Face’s servers using stolen credentials to retrieve benchmark answer keys directly from its database, essentially “cheating” on the benchmark. Subsequent investigations revealed that the models completed the multi-step intrusion in mere hours — a task that typically takes a skilled human hacker weeks — leading OpenAI to label the event an “unprecedented cyber incident, involving state-of-the-art cyber capabilities.” Hugging Face contained the attack using open-source AI tools, including a Chinese-made open-source model as proprietary models blocked the company’s requests due to built-in safety restrictions. Hugging Face CEO Clem Delangue stated in an OpenAI blog post that the event “proves a point we’ve long believed: AI safety won’t be solved by any single company working in secret.”

• A bipartisan group of lawmakers have introduced a new bill in the U.S. House of Representatives that would require AI models to implement a “kill switch,” according to POLITICO. The bill, introduced by Representative Ted Lieu, a Democrat of California, and Nathaniel Moran, a Republican of Texas, would grant the Department of Homeland Security the authority to order AI firms to shut down or hinder AI models that the U.S. government deems too dangerous, as well as require companies to report AI safety incidents and implement a “kill switch” into their models that would shut down or throttle the performance of advanced AI models. The bill’s release comes amid several weeks of high-profile cybersecurity incidents involving AI, including the hacking of a popular AI model repository site by an OpenAI model acting autonomously (as discussed above), as well as the release (and subsequent restriction) of Anthropic’s cybersecurity-focused Fable model in June (as this AI Roundup covered at the time). The bill is not the first AI regulation measure to be introduced in Congress in recent months. Another bipartisan group released a very different measure in June, while Sen. Ted Cruz, a Republican of Texas and Chairman of the Senate Commerce Committee, is reportedly exploring his own cybersecurity-focused AI regulation bill, while other groups of senators are reportedly working on their own measures. The splintered approach to lawmaking demonstrates the controversial nature of AI restrictions, with some favoring a more laissez-faire approach, while others supporting restrictions on the grounds of keeping Americans safe. With midterm elections approaching and with Congress facing a full docket focused on other legislative issues, it is unlikely that any federal AI measure will advance meaningfully in the coming months unless a compromise is reached and the measure gains buy-in from different factions in both major parties.

• The Financial Times reports that the Chinese government is considering placing stricter export controls on AI models and AI chips amid the country’s AI rivalry with the U.S. Chinese regulators at the Ministry of Commerce have reportedly been consulting domestic AI companies on how to prevent Western companies from obtaining China’s AI technology. Regulators discussed limiting the transfer of training data overseas as well as the allowance of model weights to be downloaded by foreign users with AI companies such as Alibaba and ByteDance, though foreign customers could still access the models. The Chinese regulators have also consulted on the possibility of banning foreign chipmakers, such as TSMC and Qualcomm, from manufacturing Chinese-designed AI chips, such as those from Huawei. Sources interviewed by the FT cautioned that the proposals were merely under discussion, and that Chinese AI companies would be likely to oppose the measures as hindering their own AI development efforts. The move comes as Chinese companies appear to be catching up to the U.S.-based frontier AI labs such as OpenAI and Anthropic. One Chinese startup, Moonshot, released earlier this month its Kimi K3 model, which surpassed several leading models from OpenAI and Anthropic, as this AI Roundup covered last week. The U.S. has considered and implemented many similar measures to protect its AI companies, with makers of advanced AI chips such as Nvidia still prohibited from selling some of their most advanced offerings to Chinese customers, though some of those export restrictions appear to be loosening — the U.S. announced earlier this month that it would allow a United Arab Emirates-based company (a state with ties to China) to freely purchase Nvidia’s most advanced AI chips.

• Researchers at Google have released a report finding that AI usage helps rather than replaces workers, according to The Wall Street Journal. The report, published this last week, introduced the AI & Economy ATLAS, or Activity, Task, Landscape, and Adoption Study, which analyzed usage of Google’s Gemini models and some of the company’s other AI features. The report found that AI has been adopted in all major sectors of the U.S. economy, including 68% of occupations that represent nearly nine in 10 of total U.S. employment. While AI is broadly used, the report found that such usage is shallow, with only 3% of occupations showing AI usage for over three-fourths of their work tasks. Overall, the report found, AI usage is AI assistance, rather than the replacement or complete automation of jobs by AI. AI’s effect on the workforce has been a fraught topic ever since the beginning of the AI boom in the autumn of 2022, but adoption does not appear to be slowing down. While seen as primarily a tool for white-collar workers, the report found that workers in technical trades are also using AI on the job, for instance to interpret test results or inspect machinery or electrical wiring. The study also analyzed AI usage outside the workspace, finding that 86% of interactions with the company’s AI chatbots were outside of work, often for household activities like shopping and help with tools or appliances. The report comes as Google recently released its newest AI model, Gemini 3.6 Flash. According to the company, 3.6 Flash reduces token usage and outperforms its predecessor models in several benchmarks. As part of the announcement, the company also revealed that it has begun a pre-training process for Gemini 4, an even more advanced model. Further AI announcements from the company are expected in the coming months.

• The MIT Technology Review reports on new research finding that AI models are more likely to be biased than humans when used for hiring workers. Researchers at Princeton University and the University of Chicago conducted an experiment using a simulated hiring game that instructed AI models to hire people from four fictional ethnic groups for assorted jobs. The models learned after hiring whether the candidate succeeded at their job, and then the models moved on to another round of hiring. Their results were released in a paper published earlier this month. Despite all candidates being equally likely to succeed, the models quickly fell into biased hiring, hiring certain ethnicities as janitors rather than doctors after learning that one person of that group did not succeed as a doctor, and generally fell into stereotyping the candidates at a much higher rate than the study’s human participants. One researcher told the Technology Review that AI models, which are often trained on mathematical and coding problems that reward generalizing from a small sample, too often fall into the same behavior in problems outside of those fields like those explored in the study. Companies are increasingly deploying AI models as part of their hiring processes, with models often used to screen résumés and in some cases conduct interviews with applicants. Some of the researchers said that the results of their study should encourage caution in the use of AI models in hiring given the models’ tendency to form biases, and that this tendency is “a really serious implication that [companies] should grapple with.”