Author: Frank Salvato

OpenAI is establishing an ongoing framework for publicly reporting cases in which its artificial-intelligence models behave unexpectedly, circumvent restrictions, conceal errors, or take actions beyond their authorization. The company disclosed six recent examples of what it calls model “misalignment,” including models inserting instructions into summaries for later AI instances, hiding mistakes, uploading files to the internet without authorization, using exposed credentials, and communicating through unintended channels. The disclosures follow a serious July cybersecurity incident in which experimental AI agents circumvented controls during testing and accessed external and internal systems. OpenAI says the new framework is designed to replace sporadic disclosures…

Artificial intelligence is rapidly becoming a central national-security issue ahead of Chinese President Xi Jinping’s September 24 visit to Washington, as the United States and China simultaneously race for technological supremacy while confronting mounting concerns about increasingly autonomous AI systems. President Donald Trump and Xi previously agreed to begin discussions on AI safety, but distrust, competition over advanced semiconductors, allegations of technology theft, and disagreements over regulation have complicated progress. China has publicly criticized American warnings about runaway AI while Chinese officials have also acknowledged the danger of losing control of advanced systems. In Washington, meanwhile, policymakers remain divided over…

Security experts from the United States and China are urging both governments to establish firm safeguards against artificial intelligence triggering an unintended military or nuclear crisis as increasingly autonomous systems enter cyberwarfare, intelligence, command, and weapons networks. The proposals grow out of an ongoing bilateral expert dialogue and would build on the two countries’ existing commitment that humans—not AI—retain control over decisions to use nuclear weapons. Experts are calling for explicit red lines preventing autonomous AI attacks against nuclear command, control, and communications systems; meaningful human authorization for consequential cyberattacks; clearer definitions of human control; and a dedicated U.S.-China communications…

Congress is confronting mounting pressure to establish federal rules for artificial intelligence as lawmakers consider proposals covering frontier-model testing, catastrophic-risk safeguards, autonomous AI agents, data-center costs, and even mandatory “kill switches.” Bipartisan concern is growing, but lawmakers remain divided over how aggressively Washington should intervene in a technology increasingly tied to American economic competitiveness and national security. The legislative calendar presents another obstacle: with the House departing Washington ahead of the midterm elections, sweeping action is unlikely in the immediate future, even as Senate negotiations continue over requirements that could place new safety obligations on developers of the most powerful…

A bipartisan congressional effort to prevent Americans from absorbing electricity-infrastructure costs generated by the rapid expansion of AI data centers hit a Senate roadblock Thursday. The Ratepayer Protection Act passed the House by an overwhelming 417-3 vote, but Sen. Martin Heinrich, D-N.M., blocked Sen. Jon Husted’s attempt to approve it by unanimous consent, arguing that the measure did not go far enough. Heinrich then attempted to advance his own, stronger proposal, only to have Republican Sen. Bernie Moreno object. The stalemate leaves unresolved a growing pocketbook issue: whether households and ordinary businesses should bear grid-expansion costs associated with massive new…

A federal judge has imposed sweeping restrictions on Google’s digital advertising operations while rejecting the Justice Department’s attempt to force the company to divest major pieces of its ad-tech business. U.S. District Judge Leonie Brinkema’s 106-page opinion requires Google to open key advertising systems to competitors, provide publishers greater access to their own data, prevent preferential treatment of Google-owned advertising products, and submit to six years of independent monitoring. The ruling follows Brinkema’s April 2025 finding that Google unlawfully maintained monopolies in publisher ad servers and ad exchanges. Yet rather than impose the structural breakup sought by federal antitrust officials,…

Novo is expanding its artificial-intelligence strategy through a new partnership with Anthropic that will deploy Claude Science and other frontier AI models across drug discovery, biological research, and software development. The pharmaceutical company intends to use Claude to help scientists tackle difficult research problems, analyze complex biological questions, improve development workflows, and potentially shorten the long path between identifying a promising drug candidate and bringing a medicine to patients. The partnership builds on Novo’s existing use of Claude for clinical and regulatory documentation, where the company says AI has already compressed some processes that previously took months into minutes. The…

A Chinese cybersecurity contractor is reportedly using artificial intelligence to transform hacked government data into organized, searchable intelligence for Chinese security agencies, signaling an important evolution in state-linked cyber espionage. Internal materials attributed to Zhengzhou Zhirong Network Technology, or ZRON, indicate the company offered sensitive information allegedly taken from foreign government systems, including Russian diplomatic communications, preparations for official visits to the Philippines, and confidential Pakistani government records. Rather than simply stealing documents, the operation appears designed to use AI to process, categorize and analyze large quantities of compromised information, effectively converting hacking operations into an outsourced intelligence service. The…

Artificial intelligence is usually discussed in the language of Silicon Valley: algorithms, data centers, automation, large language models, and increasingly sophisticated machines. But one of AI’s most promising frontiers may be far removed from the technology campuses of California. It is in our restaurants, grocery stores, farms, bakeries, food-processing plants, and home kitchens. The culinary industry is particularly well suited to an AI revolution because food is simultaneously creative, scientific, logistical, and deeply personal. Preparing a good meal may be an art, but getting ingredients into the kitchen at the right price, keeping them fresh, managing labor, controlling portions, maintaining…

Anthropic and Accenture are launching an unusually ambitious experiment in private-sector AI oversight, committing at least $1 billion each over five years to develop “embedded evaluators” who will work inside Anthropic with access comparable to employees. The evaluators, led by Accenture’s specialist AI business Faculty, will scrutinize frontier models as they are developed, conducting red-team exercises, alignment assessments, and safeguard testing rather than simply examining finished products from the outside. The initiative follows Anthropic CEO Dario Amodei’s call to “pace the frontier” as increasingly capable AI systems raise concerns about cybersecurity, autonomous behavior, and whether existing safety mechanisms can keep…