Hackers invade Claude Mythos: The AI model that Anthropic considered 'too dangerous to launch'

In the last week, a group of hackers managed to misaccess Claude Mythos, Anthropic's artificial intelligence model that the company itself classified as too dangerous to be released to the public. The invasion was confirmed by the company, which now investigates how the invaders gained access to a system that should be isolated. The case exposes weaknesses in the security of border AI models and rekindles the debate on sector regulation.

What is Claude Mythos and why was AI considered dangerous?

Mythos is a state-of-the-art AI model developed by Anthropic, the same creator of Claude (public version). Unlike commercial models, Mythos was considered so powerful that the company chose not to launch it publicly, fearing it could be used for sophisticated cyber attacks, large-scale misinformation generation or other malicious applications. Access was restricted to a small internal team of trusted researchers and partners.

Previously, Mythos was used internally by Mozilla to fix 271 bugs in Firefox, as reported in Mozilla Correct 271 Bugs in Firefox with Anthropic AI: The Power of Mythos in Cybersecurity, demonstrating its potential for both security and risk.

How did hackers hack into the IA model Claude Mythos?

According to Anthropic itself, cited by Euronews, an unidentified group of users obtained unauthorized access to the systems hosting Mythos. The company did not detail the method used — whether it was exploitation of vulnerabilities, social engineering or internal failure — but stated that it is conducting a thorough investigation. To date, there is no information on the amount of data accessed or whether the model was copied.

Risks of AI invasion: What does Mythos leak mean?

If hackers really have access to the full model, the consequences can be serious. Mythos is designed with advanced reasoning capabilities and code generation, which makes it a potential tool to automate cyber attacks, create custom phishing, or develop undetectable malware. Just because of these risks that Anthropic decided not to release it publicly — a decision that now appears to be correct but insufficient to prevent the leak.

Anthropic actions after the AI invasion at Claude Mythos

The company stated that it is investigating the incident and taking steps to contain the leak, including the rotation of credentials and auditing of access logs. It was not disclosed if the authorities were triggered. The case also raises questions about the safety of other border AI models maintained by companies such as OpenAI, Google DeepMind and Meta.

Frequently Asked Questions about the invasion of AI in Claude Mythos

Was Mythos really as dangerous as he claimed?

Yes, Anthropic itself had classified the model as high risk, choosing not to release it publicly to avoid malicious uses in cybersecurity.

Did the hackers get full access to the model?

It's not clear yet. The company did not disclose the extent of the access, it only confirmed that there was an unauthorized invasion of the systems that house Mythos.

Can this incident affect available AI models like Claude?

For now, there is no evidence that other Anthropic models, such as public Claude, have been compromised. The investigation is ongoing.