OpenAI vs. Mythos — a rumor
A report says OpenAI may be preparing a model to rival Anthropic’s unreleased Claude Mythos, suggesting vendors are racing not just on capability but on how to package and gate dangerous features. The story is based on press reports and should be treated as unconfirmed until vendors disclose details. (indiatoday.in)
OpenAI is reportedly preparing a new model with stronger hacking skills and plans to release it first to only a small set of companies, not the general public. The report came from Axios on April 9, 2026, and OpenAI has not publicly confirmed the model or the rollout plan. (axios.com) That rumor landed two days after Anthropic announced Claude Mythos Preview on April 7, 2026 and said it would limit access because the model is unusually good at finding software weaknesses. Anthropic said partners including Apple, Google, Microsoft, Nvidia, Amazon Web Services, CrowdStrike, and Palo Alto Networks would use it for defensive work under a program called Project Glasswing. (cnbc.com) This is not the usual “who has the smartest chatbot” race. This is a race over dual-use models, which are systems that can help defenders patch holes and also help attackers find the same holes faster. (openai.com) Anthropic has been building public language for this kind of risk for almost a year. On May 22, 2025, it said it was activating Artificial Intelligence Safety Level 3 protections for Claude Opus 4, adding tighter deployment controls and stronger security around model weights, which are the stored parameters that make a model useful in the first place. (anthropic.com) OpenAI has been moving in the same direction. On February 5, 2026, it launched Trusted Access for Cyber, an identity-based pilot that gives approved defenders access to stronger cyber capabilities and backs the effort with $10 million in application programming interface credits. (openai.com) OpenAI tied that program to GPT-5.3-Codex, which it described on February 5, 2026 as its most capable agentic coding model. “Agentic” here means the model can work through longer tasks with tools over hours or days instead of just answering one prompt at a time. (openai.com) That matters because the line between “coding model” and “cyber model” is getting thin. A system that can read a huge codebase, test assumptions, run tools, and propose fixes is only a few steps away from a system that can also chain those skills into vulnerability discovery. (openai.com) The rumor around OpenAI’s next model is tangled up with another name: Spud. India Today, citing Axios and public comments from OpenAI leaders, said on April 9, 2026 that the cyber-focused model might be related to the company’s upcoming frontier model codenamed Spud, but it also said that link is still unclear. (indiatoday.in) So the real story is not that OpenAI has definitely built a Mythos clone. The real story is that both labs now appear to accept the same basic rule: if a model gets unusually good at breaking software, the launch plan starts to look less like a product release and more like controlled access to hazardous equipment. (anthropic.com, openai.com) Until OpenAI publishes a system card, a safety note, or a product page, the specifics remain rumor. What is already public is the direction of travel: fewer instant mass rollouts, more gated programs, more named safety tiers, and more big customers getting the first look. (axios.com, cnbc.com, openai.com)
