Anthropic unveiled Mythos as too dangerous to release — and the problem reaches well beyond banks

The Mythos story matters not because Anthropic scared bankers, but because it showed every business a new reality: the time between discovering a software hole and a real attack has almost vanished. Companies used to operate on the logic of "first we disclose the vulnerability, then customers calmly install the patch." With more autonomous AI models, that scheme is starting to break down. And the most vulnerable segment is not Wall Street but thousands of less well-protected organizations that have neither a staff of expensive specialists nor the budget to rebuild their defenses instantly.

When Anthropic announced Mythos on April 7, 2026, and at the same time made clear the model was too dangerous for an ordinary release, the first reaction was predictable: this is a problem primarily for banks and critical infrastructure. A few days later, US Treasury Secretary Scott Bessent gathered the heads of the largest banks to make sure they were taking the threat seriously. For Anthropic this was ideal advertising: the company got maximum attention and at the same time raised an uncomfortable question about who gets early access to such a powerful technology at all.

According to Bloomberg, the US Treasury itself is now also seeking access to Mythos. Meanwhile, the UK's AI Security Institute already has access to the model. It has effectively become one of the main independent arbiters on AI safety. The institute's preliminary assessment boils down to one important point: there is indeed a lot of noise around Mythos, but it is not unfounded. The model is noticeably better suited to complex cyberattacks than other popular AI systems, such as ordinary chatbots like ChatGPT or Gemini. But it looks especially dangerous not against the best-protected targets, but against simplified, poorly defended systems.

This shifts the emphasis of the whole debate. Large banks typically run some of the most heavily protected IT environments in the world. Small and medium-sized businesses, regional service companies, healthcare organizations and any business where security has been treated as a secondary task for years are in a much worse position.

For a long time the industry relied on the responsible disclosure model. A vendor finds a problem, publishes a notice and offers a fix, while customers test the patch, approve the changes and only then roll them out to production. At Microsoft this has become routine, such as the monthly Patch Tuesday. In banks and large corporations the process can take weeks or months: IT teams must make sure an update will not break legacy systems, critical integrations and internal procedures. Before generative AI this was tolerable precisely because attackers usually needed even more time to work out how to turn a published bug into a working exploit.

Now that buffer is disappearing. A couple of years ago, an attacker could copy a vulnerability description into a chatbot, ask it to study public repositories and find similar patterns in other software. Today, with agentic models, the risk is higher: such systems do not merely suggest an attack idea, they can independently try variants, search for chains of weaknesses and carry an attack through to the result. According to Anthropic's description, Mythos can chain several non-fatal bugs into a multistep hacking scenario, much like a thief who first finds a slightly open window, then opens the door from inside, and then disables the alarm. No single step gives full access; together they do.

This is an important shift also because generative AI has so far mostly amplified old techniques: it helped write more convincing phishing emails, create plausible deepfakes for phone calls and speed up routine attack preparation. Agentic AI moves automation into the craft of hacking itself. And criminals have long had their own logic: they less often go head-on against the biggest banks, because that is too expensive and difficult, and more often look for hospitals, small online stores or companies with poorly configured infrastructure.

For such targets, what matters is not even whether the attacker has Mythos, but that the window between a vulnerability's publication and its exploitation has shrunk to a dangerous minimum. According to zerodayclock.com, the average time between public disclosure of a software bug and creation of a working exploit has fallen from 771 days in 2018 to less than four hours now.

What this means

Mythos is not only a striking story about a "too dangerous AI" but also a signal that the old mode of cyber defense is ceasing to work. Banks can probably adapt faster: they have people, money and processes. The main problem is everyone else. Small and medium-sized businesses will have to update systems almost in real time, and without new technical support and tougher regulatory rules, the market is unlikely to deliver that speed on its own.