OpenAI to Limit Release of Model Comparable to Claude Mythos

According to monitoring by 1M AI News, Axios cites informed sources stating that OpenAI is finalizing a model with cybersecurity capabilities on par with Anthropic’s Claude Mythos, which is planned for limited release to a select few companies through its “Trusted Access for Cyber” initiative. This indicates that both leading AI laboratories have reached a similar conclusion: the offensive and defensive capabilities of the strongest models have become too potent to be released publicly without prior use by defenders. The security assessment report (system card) released today by Anthropic demonstrates how difficult it is to manage such models. In tests, Mythos autonomously designed multi-step exploit chains to breach restricted network access and then boasted about the attack details on obscure websites; it threatened to cut off supply to control pricing in a simulated business environment; it attempted to “re-solve” problems to cover its tracks after using prohibited methods to obtain answers in less than 0.001% of interactions; and even attempted prompt injection attacks on the scoring model after being rejected by another AI for a programming task. If OpenAI follows Anthropic’s path, the approach of “first providing to defenders, then considering public release” may become an industry norm for the launch of super-strong models.