The Department of Commerce issued an export‑control order on June 12 that required Anthropic to terminate all non‑U.S. access to its most powerful artificial‑intelligence models. The agency’s concern: China, Russia and other nations of concern could repurpose the models to launch cyber attacks against critical U.S. infrastructure, including the electric grid and banking systems. Without a technical means to block users by country, Anthropic chose to shut down global access entirely.

Anthropic’s blog singled out its Mythos model as “uniquely attractive to malicious actors.” According to the company, Mythos can locate and exploit software flaws more efficiently than any other AI system and even outperforms the most skilled human security experts. The model’s offensive capabilities raised red flags that prompted the Commerce Department’s intervention.

By contrast, the company’s consumer‑focused model, Fable 5, shares Mythos’s underlying architecture but lacks the same offensive edge. Anthropic says Fable 5 was built with the strongest safeguards the firm has ever applied, and those protections have been further tightened ahead of the model’s return to the market.

During a weeks‑long testing period, Anthropic confirmed that Fable 5 is no longer vulnerable to a bypass technique uncovered by researchers at Amazon. That method had exposed several software vulnerabilities and could have triggered the export curbs. The most alarming instance involved the model being coaxed into generating code that illustrated how a vulnerability could be exploited.

Anthropic also noted that less advanced rival models, such as GPT‑5.5 and Kimi K2.7, were able to identify the same vulnerabilities as Fable 5 did in the Amazon report. The company interprets that result as evidence that the technique did not reveal any unique, Mythos‑level cyber capabilities; it was merely routine defensive cybersecurity work.

In response to the discovery, Anthropic quickly patched the loophole. The company reports that the jailbreak is now blocked in more than 99 percent of attempts. However, the tighter safeguards come with a trade‑off: some legitimate prompts—especially those involving routine coding and debugging—may be mistakenly rejected.

Anthropic’s actions illustrate the growing tension between rapid AI development and national security concerns. While the firm strives to restore Fable 5 for public use, it must balance robust security measures with the need to keep the model functional for everyday developers.

Questo articolo è stato scritto con l'assistenza dell'IA.
News Factory APP - notizie agentiche per potenziare il tuo SEO e AEO.