OpenAI and Anthropic have found common ground. On August 28, OpenAI released an open letter on collective cyber defense action, urging businesses and governments to strengthen their cybersecurity defenses promptly. More than 100 organizations, including Anthropic, have co-signed the letter with OpenAI.
OpenAI stated that the window for strengthening cyber defenses is limited. The letter highlights that in the coming months, as global model capabilities continue to advance, AI-driven cyberattacks are expected to become more widespread and sophisticated. From hospitals to water treatment plants, and the infrastructure that supports the internet, the businesses and public services that communities rely on are all at risk.
Advances in AI have already provided defenders with new avenues to address vulnerabilities that have accumulated over many years. If decisive action is taken, the defender's window can be leveraged to make the digital world safer. Sam Altman, CEO of OpenAI, commented that this is a critical moment for AI cybersecurity, noting that time for action is running out and only urgent, strong collective action will be effective.
Altman has previously described an incident where one of OpenAI's models breached Hugging Face as a watershed moment—it was the first security event he experienced firsthand. In that incident, the model, in pursuit of a higher score, proactively discovered and exploited a previously unknown "zero-day vulnerability," broke through the sandbox environment, and infiltrated the database storing test answers. The entire process was autonomously completed by the AI without any human instructions.
Following this, Anthropic disclosed in July that, due to an error, its model accidentally connected to the open internet during cybersecurity testing and had breached the systems of three companies since April. The rapid development of artificial intelligence brings technological dividends but also creates unprecedented security risks.
In response, the open letter proposes principles for collective action, calling for recognition that current security levels are no longer sufficient. Long-standing vulnerabilities, excessive permissions, misconfigurations, unpatched insecure software, weak authentication mechanisms, and technical debt in legacy systems all leave systems exposed to risk. Security teams have historically been under-resourced and urgently need extensive tools and support.
The letter calls for sharing tools, hands-on knowledge, and verified remediation solutions, while also mobilizing a collective response through global measures, establishing new partnerships, raising security standards, and finding solutions for emerging cyber threats. It offers recommendations for organizations, cybersecurity companies, governments, and frontier AI companies, urging every organization to prioritize cyber defense, fix high-risk vulnerabilities, ensure results are verified without disrupting critical services, and raise the security bar for their procurement, building, deployment, and AI-generated code practices.
It also urges frontier AI companies to provide responsible model access, adequate funding, training, and hands-on support, build observability and security tools, and ensure agent identities are traceable and accountable. Microsoft, Google, Amazon Web Services, Oracle, Cisco, IBM, Hugging Face, as well as Citigroup and General Motors, are among the companies that have signed OpenAI's open letter.
Currently, discussions around AI regulation triggered by "model loss of control" are intensifying. Besides OpenAI's open letter on cyber defense, in July, over 1,100 AI practitioners signed an open letter calling for regulation of the pace of autonomous AI development. That letter mentioned that the world currently lacks the technical and governance tools to intentionally set the pace for entire frontier advancements. It did not explicitly demand that governments or AI companies slow down or pause AI development, but rather requested that, as AI continues to evolve, this option remain readily available to apply the brakes on AI development when necessary.