Five Democratic US senators asked OpenAI CEO Sam Altman to explain the company’s AI safety, cybersecurity and employment practices in July 2024. Their letter followed reports that raised questions about whether OpenAI’s internal arrangements matched its public commitment to safe and responsible development.
Brian Schatz, Ben Ray Luján, Peter Welch, Mark Warner and Angus King requested answers by 13 August. Their questions covered computing resources for safety, employee agreements, whistleblower channels, model security, independent testing and the company’s commitments to the US government.
The inquiry connected corporate governance with national security. OpenAI was already working with government and security agencies on cybersecurity tools, making the protection and evaluation of its models a matter of public concern rather than only an internal company decision.
The senators’ letter contained twelve questions and demanded documentation rather than general reassurance. Its deadline created a specific opportunity for OpenAI to show how voluntary promises were reflected in day-to-day policies.
It also gave Congress a basis for judging the company’s response.
Lawmakers Asked Whether Safety Received Promised Resources
The senators requested information about OpenAI’s commitment to dedicate 20% of its computing resources to AI safety research. The question sought a measurable account of how the company translated a public safety promise into infrastructure and research capacity.
Compute allocation mattered because advanced safety work could compete with product development for the same scarce chips and engineering resources. A percentage commitment provided a clearer test than a general statement of intent, but only if the company documented what counted as safety research and how the allocation was calculated.
The lawmakers also asked about post-release monitoring and lessons learned from deployed systems. They wanted OpenAI to publish retrospective impact assessments, creating a record of how models behaved after launch and what the company changed in response.
Together, the questions covered the complete model lifecycle: resources devoted to research, testing before release and monitoring afterwards. This framing treated safety as a continuing operational responsibility rather than a single evaluation at launch.
Employment Agreements Came Under Scrutiny
The letter examined non-disparagement provisions affecting current and former employees. Senators asked whether OpenAI would remove contractual terms that could be used to penalise people for publicly raising concerns about company practices.
They also sought details about internal processes for reporting cybersecurity and safety issues. A credible channel needed to allow employees to escalate serious concerns without retaliation, particularly when those employees had access to technical information unavailable to the public.
Employee protection was therefore linked directly to technical oversight. If specialists feared contractual or professional consequences, important warnings might not reach leaders, regulators or the public until after a system had been deployed.
OpenAI’s Supplier Code of Conduct included non-retaliation policies and whistleblower channels, and the senators asked whether the company followed those standards itself. The comparison highlighted a governance question: whether protections expected from suppliers also applied within the organisation setting the rules.
Model Security and Independent Testing Were Priorities
The senators requested an explanation of security controls intended to prevent the theft of models, research and intellectual property. The issue was especially sensitive because a stolen advanced model could be copied or misused outside the safeguards maintained by its developer.
They asked OpenAI to commit to independent expert testing before releasing systems. External assessment could expose weaknesses that internal teams missed and provide a more credible basis for claims about safety.
One of the most consequential requests concerned government access. The senators wanted OpenAI to make future foundation models available to US agencies for pre-deployment testing, review and analysis. That would give public authorities an opportunity to assess risks before a powerful system reached the wider market.
Voluntary Commitments Faced a Test of Accountability
The inquiry referred to safety and security commitments made voluntarily by major AI companies to the Biden-Harris administration. The senators described those commitments as an important step towards public trust and asked OpenAI to document how it was meeting them.
Voluntary promises could move faster than legislation, but their value depended on transparency and verification. Questions about compute, testing, monitoring and employment protections turned broad principles into evidence that could be checked.
The letter also arrived amid reports of internal disagreements over safety and alleged cybersecurity breaches at OpenAI. The senators did not treat those reports as final findings; they used them as grounds to seek a detailed response from company leadership.
Their wider argument was that vulnerable AI systems could affect both economic competitiveness and national security. As developers partnered with government, the boundary between private product decisions and public oversight became less distinct.
OpenAI’s answers had the potential to shape expectations for the industry. If a leading model provider accepted independent testing, protected internal critics and published retrospective assessments, lawmakers could use those practices as reference points for other companies.
The inquiry therefore mattered beyond one letter. It represented an attempt to convert the language of responsible AI into specific obligations involving resources, contracts, security and external review. For OpenAI, the challenge was to demonstrate that its internal operations supported the safety commitments it made publicly.
