
Claude Mythos: what an AI model that finds software flaws means for your security
What Claude Mythos is, who can use it, what the June suspension of access taught us and what a mid-sized company should change in its security.
Guide
This guide is for owners, CEOs and directors who suspect that employees are using ChatGPT or other AI assistants on private accounts – and do not know whether that is a problem. We explain exactly what the risk is, how to gauge the scale in a few weeks and what to replace private tools with. We cover the broader topic of rules for using AI in a separate article.

Shadow AI means using AI outside the company’s knowledge and control. Take a hypothetical manufacturing company with 150 employees that has not given its staff any AI tool. A salesperson pastes a price list with discounts into a private ChatGPT account to write a quote faster. Someone in HR asks another tool to compare two CVs, and the purchasing manager asks for a summary of a supplier contract.
None of them wants to harm the company. The tool is available in a browser in five minutes, it helps them finish work faster, and nobody has told them what is not allowed. It resembles the private Dropbox accounts people once used to send files – except that now the content of documents goes into the service.
The scale is large. In the 2024 Work Trend Index survey, three quarters of knowledge workers said they use AI at work. Of these, 78% use tools not provided by their employer.
The risk does not come from ChatGPT or Claude being “dangerous”. It comes from the terms on which a private account operates. A consumer service is a contract between the provider and the employee, not the company. The company has no visibility of what went into the tool and cannot delete it when the employee leaves.
We list the specific consequences below. The table at the end of this section compares private and business versions of the same tools. The difference lies in the contract, model training and administrator control, not in the quality of the answers.
| Issue | Employee’s private account | Business version (e.g. Copilot Chat, ChatGPT Business, Claude Team) |
|---|---|---|
| Model training on conversations | Depends on the settings chosen by the employee | Off by default under the provider’s terms |
| Data processing agreement with the company | None – the employee is the party to the contract | Yes, the company is a party to the contract |
| Administrator control | None – the company cannot see accounts or conversations | The company manages accounts and can block them when an employee leaves |
| Sign-in | Private email | Work account, e.g. Microsoft Entra ID |
| Activity trail | Only in the employee’s private history | Available to the company, to an extent that depends on the plan |

Before you block or buy anything, find out who uses AI, for what and with what data. Two to three weeks and three sources of information are enough. The first is the most important, because it also tells you what people need.
Present the survey as fact-finding, not a search for culprits. If the company’s first reaction is punishment, people will stop telling the truth and will keep using AI – just more carefully.
The effective answer to shadow AI is a company tool that is at least as convenient as a private account. If your company works on Microsoft 365, the first step may cost nothing. Microsoft Copilot Chat is available at no extra cost to users with a work account and has the same contractual protections as email and files in Microsoft 365.
The no-extra-cost version of Copilot Chat works on what the employee pastes in or attaches, and on web results. For the assistant to search company documents, emails and meetings by itself, you need a paid licence. According to Microsoft’s price list (annual payment, October 2026), Microsoft 365 Copilot Business for companies with up to 300 users costs USD 21 per month, or USD 18 on promotion until 31 December 2026. Microsoft 365 Copilot with no user limit costs USD 30.
Other tools also have business versions: ChatGPT Business and Enterprise, and Claude Team and Enterprise. Choose one main tool for the whole company, not one for each department. Before switching on paid Copilot, review permissions in SharePoint and Teams. The assistant does not learn from company data, but it sees everything a given user has access to – including folders mistakenly opened to the whole company.

Shadow AI does not come from employees’ bad intentions, but from the company not giving them a safe route to a tool they need anyway.
We describe a full classification of data and permitted tools in our article on rules for using AI. In day-to-day work, however, employees need something simpler. Three questions are enough, and they are worth printing and putting up by desks, or pinning in a Teams channel:
A clear message works best: the company wants you to use AI, but in this specific tool and under these rules. Announce a transition period in which employees move their work off private accounts without consequences. Ask them to delete conversations containing company data from their private history at the same time.
A short training session with examples from your own departments’ work achieves more than a general e-learning course on artificial intelligence. The attendance list and training materials are also evidence that the company is meeting Article 4 of the AI Act, i.e. taking steps to build employees’ knowledge of AI. Turn the uses gathered in the survey into a list of ready-made examples: “how to prepare a quote”, “how to summarise a contract”.
Add one point to the procedure for employees leaving the company. When equipment is returned, ask about private AI accounts used for work and ask for company content to be deleted.
For most tasks – writing emails, summarising documents, help with data analysis – a business version of a popular tool and simple rules are enough. There are, however, situations where you need to go further. These are most often three cases:
In a mid-sized company, getting shadow AI under control usually takes a few weeks of work by one responsible person, supported by IT and a lawyer. The most common mistakes are a ban with no alternative, buying licences without rules, and blocking websites without explaining why. Each ends the same way: employees move to their private phones, and the company sees even less than before.
The order matters: first fact-finding, then the tool, and finally the rules and training. The steps worth starting with:
Related topics in the knowledge base
FAQ
The questions management boards ask most often when they discover that employees are already using AI.
About this page
Sources
About the author
Kacper Włodarczyk
CEO of ALGORCOMP
Kacper specializes in implementing Microsoft 365 Copilot, Copilot Studio, Power Platform (Power Automate, Power Apps, SharePoint) and AI agents for mid-sized B2B companies in Poland. He leads dozens of projects covering AI strategy, Power Platform governance, document workflow automation and sales processes. In his writing he focuses on the practical side of AI adoption in organizations — from the first proof of concept to company-wide rollout — with particular attention to data security, compliance (GDPR, NIS2, the AI Act) and return on investment.
Meet the teamWe will help you carry out a fact-finding exercise, choose a company tool and prepare simple rules and training for employees.
Featured

What Claude Mythos is, who can use it, what the June suspension of access taught us and what a mid-sized company should change in its security.

What a vCISO does, what Poland’s new National Cybersecurity System Act requires of companies, when this model pays off, and when a full-time hire or a good IT provider is the better choice.

A comparison of costs and models: an in-house CISO, a vCISO for an agreed number of days (a part-time CISO) and a hybrid model. How to calculate the full annual cost of each and when to choose which.