OpenAI, Anthropic and Google DeepMind have been holding discussions around AI safety for several weeks, bringing together three of the most influential companies building frontier artificial intelligence systems.
The talks were confirmed by OpenAI global policy chief Chris Lehane and come at a time when concerns around increasingly capable AI models are moving closer to the center of industry and government debate. The companies remain fierce competitors, but they are also facing the same question: how far can AI capabilities advance before stronger safeguards become unavoidable?
The discussions are reportedly focused on independent evaluations, common safety standards and possible mechanisms for monitoring advanced AI systems. That is a notable shift for an industry that has often relied on voluntary commitments and internal testing.
OpenAI Confirms Weeks of AI Safety Talks
OpenAI has confirmed that it has been speaking with Anthropic and Google DeepMind about AI safety for several weeks. The conversations involve some of the biggest names in frontier AI and show that safety concerns are no longer being handled only through public statements or individual company policies.
Lehane revealed the talks while discussing AI policy in Washington. His comments came as leading AI executives increased their warnings about the risks of systems becoming more powerful, autonomous and difficult to evaluate.
The timing also follows renewed calls from Anthropic CEO Dario Amodei for companies to consider how quickly frontier AI should advance. OpenAI CEO Sam Altman and Google DeepMind CEO Demis Hassabis have also spoken publicly about the need for stronger safeguards around advanced systems.
Anthropic Pushes for a More Controlled AI Race
Anthropic has emerged as one of the strongest voices arguing that frontier AI development needs more deliberate safeguards.
Amodei has suggested that the industry should consider ways to slow the pace of development enough to give researchers, governments and independent evaluators more time to understand the risks attached to more powerful models.
The idea is not necessarily to stop AI development. It is more about avoiding a situation where model capabilities advance faster than the safety systems designed to evaluate them.
Anthropic has also supported the use of outside evaluators who could inspect frontier models and assess whether companies are following their own safety commitments. This would reduce the industry’s current dependence on self-assessment.
Independent Evaluators Could Play a Bigger Role
One of the strongest themes emerging from the discussions is the role of independent verification.
Frontier AI companies currently conduct much of their model testing internally. They publish safety reports, conduct red-team exercises and establish their own thresholds for deployment.
Independent evaluators could introduce another layer of oversight. These organizations could examine high-risk capabilities, review safety procedures and provide external assessments before powerful models are released.
OpenAI has indicated support for this broader approach. The company has backed policy proposals that would give qualified outside organizations access to frontier systems for verification purposes.
That would represent a major change in how AI safety is handled. Instead of asking companies to simply certify that their own systems are safe, an external group could test those claims directly.
A Frontier AI Standards Body Is Being Discussed
The conversations between OpenAI, Anthropic and Google could eventually lead to a more formal industry safety organization.
The companies have reportedly discussed the possibility of creating a standards body focused on frontier AI. Such an organization could establish shared testing methods, identify dangerous capabilities and create common expectations for companies developing advanced models.
Hassabis has previously raised the idea of an independent body capable of monitoring powerful AI systems. The concept would be similar to oversight structures used in other high-risk industries, although applying that model to fast-moving AI development would be complicated.
A standards organization could also create more consistency. Right now, different companies use different safety frameworks, testing methods and deployment thresholds.
AI Rivals Face an Antitrust Question
Cooperation between major AI companies also comes with legal complications.
OpenAI, Anthropic and Google are direct competitors. Any agreement involving shared development practices, restrictions or coordinated decisions could attract attention from competition regulators.
Anthropic has suggested that governments may need to provide limited antitrust protections for companies cooperating specifically on AI safety.
The argument is that frontier laboratories should be able to share safety information or coordinate around serious risks without worrying that those conversations could later be interpreted as anti-competitive behavior.
OpenAI appears less convinced that such protections are necessary. Lehane has indicated that AI companies can already cooperate on safety issues within existing legal frameworks.
The issue may become more important if the current discussions move beyond informal conversations and turn into actual industry-wide standards.
OpenAI Supports Stronger Safety Verification
OpenAI has also been pushing for more formal safety requirements around advanced AI systems.
The company has expressed support for policy measures that would require frontier developers to provide access to independent verification organizations.
Such groups could evaluate whether companies are meeting safety requirements and whether their models contain capabilities that could create serious risks.
OpenAI has also argued that AI governance needs to become more capability-based. Instead of regulating models simply because of their size or the amount of computing power used to train them, policies could focus on what those systems are actually capable of doing.
That approach would allow regulators to focus on models that pose meaningful risks while avoiding unnecessary restrictions on less capable systems.
Washington Remains Divided Over AI Regulation
Even as AI companies discuss stronger safety measures, policymakers remain divided over how much regulation is necessary.
Some officials argue that stricter rules could slow American AI development and weaken the country’s position against China. Others believe the risks of frontier AI are serious enough to justify tougher oversight.
This tension has become central to the AI policy debate in Washington.
Companies want enough freedom to continue innovating, but they are also increasingly asking governments to create clearer rules around advanced systems.
Lawmakers, meanwhile, are still debating whether those rules should focus on voluntary standards, independent testing, mandatory reporting or stronger government oversight.
The AI Safety Debate Is Becoming More Serious
The discussions between OpenAI, Anthropic and Google show how quickly the AI safety conversation has changed.
Only a few years ago, many safety proposals were treated as long-term questions. Today, some of the companies building the most powerful models are openly discussing independent evaluations, common standards and limits around frontier development.
That does not mean the companies suddenly agree on everything.
They are still competing for customers, researchers, infrastructure and market leadership. Their incentives are not always aligned, and any industry-led safety system will face questions about transparency and accountability.
Still, the talks are significant because they suggest AI safety is moving beyond company-specific policies.
The industry is starting to explore whether frontier AI may need shared rules before models become significantly more capable than the systems available today.
What Happens Next
The biggest question is whether these discussions lead to something concrete.
An industry standards organization, independent verification system or shared safety framework would represent a substantial change in how frontier AI companies operate.
But turning discussions into enforceable rules will not be easy.
Companies would need to agree on testing standards, definitions of dangerous capability and procedures for responding when a model crosses a particular risk threshold. Governments would also have to decide whether those standards remain voluntary or become part of formal regulation.
For now, the talks show that OpenAI, Anthropic and Google are at least willing to cooperate on one issue despite competing aggressively everywhere else.
The AI race is still moving fast. The companies leading it are now spending more time discussing what could happen if it moves too fast.
Sources
TechCrunch: OpenAI, Anthropic, Google have been in talks on AI safety for weeks
https://techcrunch.com/2026/09/15/openai-anthropic-google-have-been-in-talks-on-ai-safety-for-weeks/
OpenAI: The AI policy window is open. We need to act.
https://openai.com/index/ai-policy-window/
Bloomberg Law: OpenAI Says It’s Working With Anthropic, Google on AI Safety
https://news.bloomberglaw.com/privacy-and-data-security/openai-says-its-working-with-anthropic-google-on-ai-safety
The Washington Post: Leading AI companies discussed creating new safety body
https://www.washingtonpost.com/technology/2026/09/14/anthropic-openai-google-discussed-creating-new-ai-safety-body/
