Key Points:
- The U.S. government added Google DeepMind, xAI, and Microsoft to its artificial intelligence safety program.
- Federal scientists will test unreleased software to prevent cyberattacks and the creation of biological weapons.
- OpenAI allowed the government team to test its GPT-5.5-Cyber model to find hidden security flaws.
- Security researchers have already discovered and fixed dangerous vulnerabilities in tools built by Anthropic and OpenAI.
The Trump administration announced on Tuesday that it will expand a federal program that gives government scientists early access to unreleased artificial intelligence models. The initiative now includes major technology companies like Google DeepMind, xAI, and Microsoft. These technology giants join OpenAI and Anthropic, who already voluntarily agreed to work with the United States Center for AI Standards and Innovation. This federal team of scientists wants to identify hidden vulnerabilities in software before companies release their final products to the general public.
Government officials focus their attention on clear, demonstrable risks that threaten national security. Federal scientists specifically look for software flaws that hackers could exploit to launch massive cyberattacks against American infrastructure systems. They also work hard to block foreign adversaries from using artificial intelligence to design chemical or biological weapons. Additionally, the federal team monitors the systems to ensure that bad actors cannot corrupt the massive datasets used to train American software models.
Each technology company plans to hand over different tools and data for the scientists to test. Chris Lehane, the head of global affairs at OpenAI, posted on LinkedIn on Tuesday that his company gave the government access to a model called GPT-5.5-Cyber. This specific software variant focuses heavily on defensive cybersecurity work. The scientists will test the model to see exactly how it handles intense digital threats from outside attackers.
Microsoft plans to take a slightly different approach to the partnership. The software company stated that it will work directly with government scientists to build shared datasets and new workflows. These tools will help federal workers effectively assess advanced artificial intelligence models across the industry. However, Microsoft did not name exactly which of its own software models it will submit for testing.
Meanwhile, Anthropic gave the government agency deep access to both its publicly available products and its secret, unreleased models. The company allowed federal researchers to probe the software for weak spots through a rigorous process known as red-teaming. During this process, government workers simulate the exact behavior of malicious hackers to see how the system reacts under pressure. Anthropic also handed over detailed internal documents that outline known safety issues and built-in security mechanisms.
Google DeepMind, the dedicated artificial intelligence research branch of Alphabet, agreed to provide federal workers with direct access to its proprietary models. A company spokesperson confirmed that the tech giant will share core data to help the government understand how the software operates behind closed doors. xAI also joined the group to help the federal government assess new technology, though specific details regarding its exact contributions remain private.
The testing program has already produced serious results and forced companies to fix dangerous flaws. During the initial tests with Anthropic, federal scientists discovered they could easily bypass the software’s built-in safety mechanisms. The researchers used simple tricks such as substituting specific characters or telling the machine that a human reviewer had already approved the request. Anthropic noted these exact flaws and quickly patched the vulnerabilities before they could cause real-world damage.
OpenAI also found major security gaps thanks to its early work with the federal agency. Back in September 2023, the government team tested a tool known as the ChatGPT Agent. The researchers found a massive vulnerability that could allow sophisticated hackers to bypass the company’s entire cybersecurity shield.
If left unfixed, this software flaw would have allowed a cybercriminal to remotely control the computer systems that the artificial intelligence agent accessed during a session. The attacker could then successfully impersonate the original user and log into their private accounts on other websites. OpenAI fixed this massive security hole immediately after the government team pointed it out.
This federal push for software safety builds upon previous agreements across the technology sector. Several major companies, including Meta, Amazon, and Inflection AI, previously agreed to allow independent experts to audit their artificial intelligence models. These early checks focused heavily on finding biosecurity threats and glaring cybersecurity risks before the technology reached ordinary consumers.
The government agency, which originally operated under a different name during former President Joe Biden’s tenure, continues to expand its mission today. The scientists recently published a set of voluntary guidelines designed to stop artificial intelligence models from leaking private health information or generating confidently incorrect answers.
Now, the federal team wants to secure the most vital parts of the American economy. According to the agency’s official website, the scientists are currently drafting new testing guidelines specifically for critical infrastructure providers. These upcoming rules will help companies in the emergency services sector and the communications industry safely test their own artificial intelligence systems before deploying them into the real world.
