Claude Mythos: what an AI model that finds software vulnerabilities means for your company’s security

Claude Mythos is the most powerful family of AI models from Anthropic, the company behind the Claude assistant. What sets it apart from earlier models is above all how well it reads program code and finds vulnerabilities in it. A vulnerability is a flaw in software that allows an attacker to take over a system or steal data.

The Firefox browser is the clearest example. Mozilla, which develops it, tested an early version of the model, Claude Mythos Preview. Firefox 150 fixed 271 vulnerabilities found during these tests, 180 of which were rated as serious.

It is worth knowing the other side too. In the same post, Mozilla noted that an outstanding specialist could also have found each of these vulnerabilities. So the model does not do anything impossible for a human. It does it much faster and more cheaply, and that changes the calculation on both sides.

The model’s history is short but turbulent. For the management board, it matters mainly because it shows how unstable access to the latest AI tools can be. We have listed the key dates in the table below.

In practice, Fable and Mythos are the same family of models with different safeguards. Fable is the generally available version. When it detects a request in a sensitive area, such as attacking systems, biology or chemistry, a weaker model takes over the response. Mythos has some of these blocks removed and is given only to vetted defenders.

The June suspension lasted less than three weeks. The US Department of Commerce ordered access to be blocked for people outside the US, and Anthropic, unable to check users’ citizenship in real time, switched both models off for everyone. When access was restored, the company agreed, among other things, to give the government early access to future models and to report any misuse it detects.

Any company with a suitable Claude plan, or through the API, can use Fable 5.1. The API is the interface for connecting the model to your own systems. In business plans, access depends on the type of plan and usage limits. In the API, the price is USD 10 per million input tokens and USD 50 per million output tokens (as of 14 October 2026); a token is roughly a fragment of a word.

Only verified organisations and security specialists get Mythos 5.1, through the Cyber Verification Program. Applications are made through an Anthropic portal, and access covers only the organisation’s own code and systems. The API price is the same as for Fable. In return for looser blocks, the programme requires, among other things, that conversation history is retained.

For a typical manufacturing or trading company that does not develop its own software, direct access to Mythos makes little sense. You will get more value from asking your software vendors and the company that runs your IT whether they use such tools themselves.

Until now, the security of many mid-sized companies has rested on an unspoken assumption: we are too small for anyone to spend time on us. Finding vulnerabilities by hand was expensive, so criminals chose big targets or ready-made, known flaws. Models such as Mythos lower the cost of this work for anyone who has access to them.

Anthropic deliberately gave the full version to defenders first, so they would have time to patch the most important software. But this advantage is temporary. Other companies and countries are developing similar models, and not all of them apply the same restrictions.

For your company, this means above all less time to react. When a vendor publishes a patch, criminals can work out more quickly which vulnerability it closes and attack those who have not installed it. A company that updates its systems once a quarter gives them more and more time to do so.

None of the changes below requires access to Mythos or a large budget. They are basics that, in a world of faster attacks, are no longer optional. They are listed in the order in which it usually makes sense to start.

Take a hypothetical distribution company with 90 employees, an ERP system, an online shop and several programs from local vendors. The first step is a list of these systems with the name of the person responsible for each. This alone usually shows that nobody is looking after updates to the online shop or an old file server.

For less than three weeks, Fable 5 and Mythos 5 were unavailable to everyone, including Polish companies. The decision was made by the government of another country, and the provider could not get around it. The same could happen with any AI tool because of regulations, sanctions, outages or a change in pricing.

The point is not to give up on AI. The point is to ask, for every important process: what will we do if this model stops working tomorrow? The answer should go into the business continuity plan, the document that describes how the company keeps working when something fails.

In practice, three rules are enough. The process should also work on a different model, even if less well. You keep the data and instructions for AI on your side, not only with the provider. People should be able to carry out the key task by hand, even if more slowly.

Since 3 April 2026, an amendment to the Polish National Cybersecurity System Act (KSC Act) has been in force, implementing the EU NIS2 Directive in Poland. It applies to essential and important entities, mainly mid-sized and large companies in the sectors listed in the Act. Companies covered by the Act have until 3 April 2027 to implement an information security management system (ISMS) and their other obligations.

The Act does not mention AI models, but it requires risk management that is proportionate to the threats. If attackers can find vulnerabilities faster, a risk assessment from two years ago may no longer be enough. The head of the entity (management board) may also be personally liable, and the fine for an essential entity can reach EUR 10 million or 2% of revenue. The authorities can impose fines from 3 April 2028, but the deadline for implementing the obligations falls a year earlier.

If your company is covered by the Act, add two points to your next risk review. The first is faster AI-assisted attacks and how long it really takes you to install patches. The second is dependence on external AI providers in important processes.

If your company still has no backups that have been tested by restoring them, no two-factor sign-in for email and no up-to-date systems on its computers, start there. Mythos changes nothing here, because criminals have been exploiting such gaps for years without any help from AI.

Likewise, there is no point in buying specialist AI tools to find vulnerabilities if you have nobody to read the results and make sure the fixes are made. A report listing a hundred problems, with nobody to fix them, does not make you any more secure. Security specialists point out that the slowest stage today is people checking and fixing flaws, not finding them.