MMOOZZGG.com

AI assistant

In short

AI now finds software vulnerabilities as well as strong security experts do, and attackers will soon use the same tools. The article covers Mozilla’s 271 fixes in Firefox 150, what an AI agent is, and how to limit its rights. Main takeaway: give an agent only the rights its own process needs.

Key ideas

  • On 7 April 2026 Anthropic presented Claude Mythos Preview and said it would not release the model publicly.
  • Mozilla fixed 271 vulnerabilities in Firefox 150 that a model helped to find.
  • Give an AI agent its own account, with rights only for its own process.
  • If personal data goes to an outside service, your company answers to clients as the data operator under Russian law.
  • The article lists five questions to ask a vendor and states that it is not legal advice.

Prepared by AI from the article

AI finds vulnerabilities at expert level: what access to give your AI agent

NewsSecurity

Advanced3 min read

AI now finds software vulnerabilities as well as strong security experts do. Mozilla fixed 271 vulnerabilities in Firefox 150 that a model helped to find. For a company, the lesson is practical: give an AI agent only the rights its own process needs, because attackers will soon use the same tools.

On this page · 5 sections

What happened with Claude Mythos Preview?

On 7 April 2026, Anthropic presented Claude Mythos Preview and said it would not release the model publicly. By the company’s own statement, it was the most capable model it had built, and it finds flaws in other people’s code too well.

  • Project Glasswing gave access to 11 partners and more than 40 other organizations that maintain important software. The partners include AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. They use the model to find and fix flaws before attackers find them.
  • 21 April 2026: Mozilla reported 271 vulnerabilities fixed in Firefox 150, found with the model.
  • 9 June 2026: Anthropic opened Claude Fable 5, a version with stronger limits, to all users. Only Project Glasswing partners received Mythos 5.
  • 12 June 2026: US authorities set export controls on both models. Anthropic could not check citizenship in real time, so it paused access for everyone.
  • 30 June 2026: The controls were lifted. From 1 July, Fable 5 was available worldwide again, and Mythos 5 returned to some US organizations.

Anthropic treats vulnerability search as so strong a skill that it limits access to it.

Why does this matter if you run an AI agent?

An AI agent is a neural network that does a task step by step: it looks up data, uses programs, and checks the result. In practice, it is a program with an account in your CRM, phone system, and messengers. It receives messages from strangers around the clock.

If the agent has full rights, anyone who tricks it into following a hidden instruction can reach those rights through it. This attack is called prompt injection. Automatic vulnerability search makes such attacks cheaper for the attacker, so they become more likely.

How do you limit an agent’s rights?

Give the agent its own account with rights only for its own process. A sales agent needs to create deals. It does not need a customer database export. A booking agent only needs to see the schedule.

Our agents work this way. The system, the data, and the access rights stay with you after the project. Check four more things with any vendor:

  1. Does the agent write its actions to your CRM log?
  2. Which outside services does the message text pass through, and is there a data-processing contract with each?
  3. Does the agent obey commands inside incoming messages that change its rules? It should not. Pass disputed requests to a person.
  4. How fast can you revoke its access?

What does Russian law add?

If correspondence with personal data goes to an outside service, your company answers to clients for its actions as the data operator. If the service is foreign, you answer together with it (Article 6, Parts 5 and 6, of 152-FZ, Russia’s personal data law). To transfer data abroad you must notify Roskomnadzor in advance (Article 12).

Under Article 13.11 of the Administrative Code as amended by law 420-FZ, in force since 30 May 2025, a company’s fine for a leak of personal data starts at $36,000 for 1,000 people or more. A repeat leak costs 1–3% of revenue, not less than $240,000 and not more than $6 million. Automatic vulnerability search makes these scenarios cheaper for an attacker, and so more likely. We process personal data on servers in Russia under 152-FZ.

What five questions should you ask a vendor?

  • Which exact rights does the agent have in the CRM, ERP, and phone system, and can they be reduced?
  • Where is the conversation history and the knowledge base stored? A knowledge base is a set of selected company documents that the assistant searches for answers.
  • Which outside services does the message text pass through, and is there a contract with them?
  • What does the agent do when a customer asks it to “forget the rules” or to send a customer list?
  • Who revokes access when the contract ends, and how?

This is not legal advice. Have a lawyer review the documents, and have your administrator review the system rights.

Questions and answers

Can an AI agent cause a data leak?

Yes, if it has full rights, if messages go through outside services, or if it obeys commands from incoming messages. You can reduce each of these risks with settings.

What rights does a sales agent need?

It needs to create and update deals, write message history, and read the knowledge base. It does not need a customer database export or administrator rights.

What is Project Glasswing?

It is an Anthropic program announced on 7 April 2026. It gives the Mythos Preview model to 11 partners and more than 40 other organizations, so they can find and fix vulnerabilities in their software before attackers do.

Contact us