“Dangerous” AI models are coming no matter what

In a significant move last week, Anthropic temporarily took its latest AI models, Claude Fable 5 and Mythos 5, offline due to a U.S. government directive restricting access for foreign nationals. The company has been engaged in discussions with the White House since Friday, but as of now, no agreement has been reached to reactivate these services. Anthropic's Mythos model, released in April, has been touted for its advanced capabilities that can identify software vulnerabilities for security enhancement. However, these same features pose a potential risk, as they could be exploited by malicious actors. In a recent blog post, the company highlighted the dual-use nature of AI technologies, emphasizing that tools beneficial for cybersecurity experts and researchers could also be dangerous in the wrong hands. Initially, Anthropic rolled out a version called Mythos Preview to a select group as part of Project Glasswing. Last week, Mythos 5 was also introduced to this consortium, while Claude Fable 5 became available to the public—with restrictions on its responses regarding sensitive topics such as biology and cybersecurity. However, the Trump administration intervened, citing concerns that the safety measures in Fable 5 could be bypassed, allowing unrestricted access to Mythos 5's functionalities, which could pose a threat to national security. Experts warn that this regulatory clash may only be a temporary setback. While Anthropic is currently at the forefront of these discussions, the reality is that numerous companies and open-source developers are likely to develop AI models with similar capabilities to Mythos 5 in the near future, if they do not already exist.