U.S. Commerce deletes AI safety audits

The U.S. Commerce Department removed from its website details of an agreement under which Microsoft, Google and xAI would submit new artificial intelligence models for government security testing before public release, according to a Reuters review published May 11 and copies of the missing announcement described by other outlets. (finance.yahoo.com) The announcement had been posted on May 5, Reuters and follow-on reports said. By May 11, the link returned a page-not-found message and later redirected to the Center for AI Standards and Innovation, or CAISI, the Commerce-linked group that runs such evaluations. (economictimes.indiatimes.com) The deleted page matters because it described one of the clearest public examples of the U.S. government getting pre-release access to frontier AI systems from major developers. (nist.gov) The surviving CAISI page says the center is the government’s main contact point for testing commercial AI systems, setting up voluntary agreements with developers and leading unclassified evaluations of models that may pose national-security risks. ### Which page disappeared, and when? Reuters reported on May 11 that Commerce had removed details of its agreement with Google, xAI and Microsoft to test AI models for security vulnerabilities. Reuters said the link that had previously led to the department’s announcement was no longer available. Other reports describing the same page said it had been posted on May 5. (finance.yahoo.com) By Monday afternoon in Washington, according to reports that cited the broken link, the page showed a “Sorry, we cannot find that page” message. Later, the link redirected to CAISI’s website. (nist.gov) ### What did the deleted announcement say the companies had agreed to? The May 5 announcement said Microsoft, Google and xAI would hand over new AI models before deploying them to the public so government scientists could test them for security flaws, according to Reuters-based reports that quoted or summarized the missing page. Those reports said the testing was aimed at risks including cyberattacks, military misuse and broader national-security vulnerabilities. (finance.yahoo.com) The arrangement was described as voluntary. CAISI’s current public page says the center is tasked with establishing voluntary agreements with private-sector AI developers and evaluators, and with leading unclassified evaluations of AI capabilities that may pose risks to national security. (theoutpost.ai) ### What is CAISI, and what does it still say publicly? CAISI’s website says the center serves as industry’s primary point of contact within the U.S. government for testing and collaborative research on commercial AI systems. The page says CAISI focuses on demonstrable risks including cybersecurity, biosecurity and chemical weapons, and coordinates with agencies including the Defense Department, Energy Department, Homeland Security, the White House Office of Science and Technology Policy and the intelligence community. (energy.economictimes.indiatimes.com) A separate NIST post from May 1 shows CAISI is continuing to publish technical evaluations. In that post, CAISI said it had evaluated DeepSeek V4 Pro in April 2026 using benchmarks across cyber, software engineering, natural sciences, abstract reasoning and mathematics. (nist.gov) ### Were any audit results or red-team findings for Microsoft, Google or xAI published? No public source I could verify shows published CAISI audit-result files or red-team scorecards for Microsoft, Google or xAI models that were later deleted on May 14. The verified reporting I found says Commerce removed the webpage describing the agreements themselves, not that the department had posted and then deleted detailed model-by-model audit outcomes. (nist.gov) CAISI’s public materials do show that red-teaming and evaluation work are part of its mission. Its site says it leads unclassified evaluations of AI systems, and NIST has separately published research on AI agent security red-teaming competitions. (nist.gov) But those materials do not, in the sources reviewed, provide public audit files for Microsoft, Google or xAI frontier models. ### What remains unanswered? Commerce had removed the announcement page by May 11, but the sources reviewed do not include a public explanation from the department for why it was taken down. The materials I found also do not show whether the underlying testing arrangements changed, ended or simply became less publicly documented. (finance.yahoo.com) CAISI’s website remains live as of May 14 and still says it establishes voluntary agreements with AI developers and leads unclassified evaluations of models that may pose national-security risks. Any future public update would most likely appear on Commerce or NIST pages tied to CAISI, which continues to post news and technical evaluation write-ups there. (nist.gov) (finance.yahoo.com)