According to external sources, Anthropic resumed a global visit to Claude Fable 5 on 1 July, and will reopen less restricted Mythos 5 to a group of audited American institutions. According to the article, this change was brought to the attention of the encryption industry because 2026 was the year of a high number of hacker attacks, and AI is reducing the time between the opening of the loopholes and their use.

Fable 5 Global, Mythos 5 is still restricted

According to the article, the resumption of the visit was not about the same level of openness. Fable 5 has been re-opened to global users, and Mythos 5 has been made available to United States institutions responsible for critical infrastructure protection only through a project called Glasswing.

This distinction is important. Markets are concerned mainly about whether models with greater network security capabilities will flow directly to ordinary users. According to the article, the currently less restricted and more capable version is not open to the public.

In June of this year, researchers demonstrated a way out of prison, allowing Fable 5 to identify software loopholes and generate usage codes. Subsequently, the U.S. Department of Commerce implemented emergency export controls and Anthropic had a offline model. By the end of June, the restrictions were lifted and the model resumed its visit in early July.

The controversy is not about whether or not to attack, but more quickly.

The article argues that the controversy surrounding such models is not about whether it invented a completely new pattern of attack, but rather about whether it made the pre-existing pattern of attack faster and cheaper.

According to the text, the Mythos series model found over 10,000 high-risk and serious loopholes in restricted tests, covering over 1,000 open-source projects, including common components such as the Linux kernel. It is also mentioned that in one case, the model generated a functioning loophole in 31 minutes to exploit the PoC.

  • The loophole is exposed to a period that can be compressed from days to hours.
  • Individual operators can use lower-cost acceleration gap analysis and use to prepare
  • The risks of open codes, old dependency and unaudited silos are rising simultaneously

Why is encryption more vulnerable?

According to the article, the encryption industry is more exposed for reasons directly related to its technical structure. Smart contract codes are naturally open, and the logic of controlling a large number of assets is exposed to the chain for long periods, accessible to anyone, including the AI model that automatically analyzes the code.

Cross-chain bridges are another high-risk target. Such agreements tend to focus on collateral assets in multiple chains, with greater losses often incurred when information validation or contract logic is flawed. For the attackers, this means a higher return and a stronger scanning power.

The article mentions that the encryption industry lost more than $840 million as a result of hacker attacks in 2026, with some of the numbers even above $940 million, involving more than 120 incidents. Against this background, AI improves the efficiency of the detection of loopholes, often first impacting projects with inadequate audit coverage, slow patches and long-term use of old codes.

The defense will use the same model.

The article also notes that the same ability does not serve only the attackers. The security team could use it to more quickly audit the code, locate the defects and advance the repair, so long-term results remain controversial. A clearer change in the short term would be to guard against the speed of both sides, with slow-responding agreements under pressure.

The core judgement is that the risk should not be focused solely on whether a particular model is open, but rather should see that the relevant capabilities already exist and that the weaker model can reproduce some of these processes. For the encryption project, the real problem is not the emergence of a new type of attack, but the chain of human and time-dependent loopholes is being automated.