Former employees from OpenAI and other major technology companies used a US Senate hearing to argue that artificial intelligence is advancing faster than the security systems around it. Their central concern was not only that powerful models may create new risks, but that companies building them may not have enough internal pressure to slow down, test carefully and invite outside scrutiny.
William Saunders, a former OpenAI employee, delivered one of the sharpest warnings. He argued that OpenAI had prioritized rapid AI development over stronger security practices, and he linked that concern to the possibility that artificial general intelligence, or AGI, could arrive sooner than many people expect.
A three-year AGI warning
Saunders told the Senate that an AGI system could be developed in as little as three years. In his view, recent performance by OpenAI's o1 model in math and coding competitions shows how quickly capabilities can move from weak to highly competitive.
He described that leap in direct terms: "OpenAI’s new system leaps from failing to qualify to winning a gold medal, doing better than me in an area relevant to my own job. There are still significant gaps to close but I believe it is plausible that an AGI system could be built in as little as three years."
The point of that warning was not simply that AI could become more capable. Saunders said a system able to perform most economically valuable work better than humans would raise serious security concerns. The risks named at the hearing included autonomous cyberattacks and assistance in the development of biological weapons.
He also gave a concrete example of why testing matters. Saunders said a new AI system from OpenAI could help experts plan the reproduction of a known biological threat. His warning was direct: "Without rigorous testing, developers might miss this kind of dangerous capability,"
Security concerns inside AI companies
Saunders also criticized OpenAI's internal controls. He said that during his time at the company, there were periods when vulnerabilities could have allowed him or hundreds of other engineers to bypass access controls and steal the company's most advanced AI systems, including GPT-4.
That allegation goes to the heart of the security debate around frontier AI. If the most advanced systems can be copied, leaked or misused, then the risk is not limited to what a company chooses to release. It also includes what could happen if internal access systems fail.
OpenAI rejected the accusations. The company said it had taken "no shortcuts in our safety process" and had conducted "extensive internal and external" testing to meet political obligations.
The criticism also arrived after other concerns about OpenAI's safety work. According to a report in the Washington Post, security tests for the AI model GPT-4 Omni were completed in just one week, which reportedly caused displeasure among some employees.
Former insiders call for rules
Saunders was not the only former insider to question whether AI companies can be left to manage these risks on their own. Former OpenAI board member Helen Toner criticized what she described as fragile internal control mechanisms. She reported cases where safety concerns were ignored so products could be brought to market more quickly.
Toner also argued that companies cannot fully consider the interests of the general public if they alone are responsible for detailed decisions on security measures. That position shifts the debate from company policy to public oversight.
David Evan Harris, a former Meta employee, warned against relying on voluntary self-regulation. He criticized reductions in security teams across the industry and called for binding legal rules.
Margaret Mitchell, who previously worked on ethical AI issues at Google, focused on incentives inside AI companies. According to Mitchell, employees who work on safety and ethics are less likely to be promoted. That creates a structural problem: the people raising risk concerns may have less career support than those pushing products forward.
What stronger AI regulation could include
The experts at the hearing called for stronger government regulation of the AI industry. Their proposals were not limited to one company or one model. They described a broader oversight system for high-risk AI systems and the organizations building them.
The measures discussed included:
- Mandatory transparency requirements for high-risk AI systems.
- More research investment in AI safety.
- A stronger ecosystem for independent audits.
- Better protection for whistleblowers.
- More technical expertise inside government agencies.
- Clearer liability for AI-related damages.
The witnesses argued that regulation would not block innovation. Their position was that clear rules could strengthen consumer confidence and give companies planning security.
Senator Richard Blumenthal, Chairman of the subcommittee, said he would soon present a draft bill on AI regulation. That makes the hearing part of a larger policy debate over whether governments should set binding rules before the most powerful AI systems arrive.
OpenAI's safety moves
The hearing also came as OpenAI was making its own safety-related changes. Since November last year, OpenAI has lost several employees working on AI safety, including former Chief Scientist Ilya Sutskever and Jan Leike, who jointly led the Superalignment Team.
This week, OpenAI introduced a new "Safety and Security Committee" led by Zico Kolter. According to the source article, the committee has broad powers to monitor safety measures in the development and introduction of AI models.
A few weeks earlier, OpenAI reached an agreement with the US National Institute of Standards and Technology (NIST). Under that agreement, the US AI Safety Institute gets access to new AI models before and after publication to work on AI safety research, testing and evaluation.
The tension is clear. Former insiders say voluntary systems are not enough, while OpenAI points to internal and external testing, a new committee and cooperation with government safety institutions. The Senate hearing framed the question now facing policymakers: how much oversight is needed before AGI-level systems, if they arrive, create risks that are harder to contain.