When new models play into the hands of cyber criminals
Anthropic has unveiled two new «Mythos-class» AI models, Fable 5 and Mythos 5, which are set to set new standards in the field of coding. However, the announcement also raises serious security issues. Oliver Keizers from Filigran explains why companies need to act now.
With the announcement of Fable 5 and Mythos 5, Anthropic has unveiled two new AI models of the so-called «Mythos class». The models promise new performance peaks, particularly in the area of coding, and have built-in security locks in high-risk areas. What sounds like a technological advance at first glance harbors considerable risks for companies and their IT security, according to Oliver Keizers, VP Sales Central EMEA at cybersecurity company Filigran.
The same technology that protects can also attack
Keizers warns of a structural dilemma: «With the announcement of Fable 5 and Mythos 5, Anthropic is sending a clear signal: AI is accelerating productivity more than ever, but unfortunately also opening up a whole new range of unknown attack vectors for cybercriminals. Once again, AI is becoming a double-edged sword. Because the same model class that helps defenders find vulnerabilities and secure software also shortens the time it takes attackers to find security holes, adapt their techniques and scale their sabotage.»
This assessment is particularly relevant for SMEs: Those with limited IT resources rely on a fast and structured response capability - precisely the ability that attackers can increasingly use against companies with the help of powerful AI models.
Guardrails alone are not enough
According to Keizers, Anthropic's integrated security barriers - so-called guardrails - are a necessary step, but not a sufficient measure: «Guardrails are necessary, but are not a complete defense strategy on their own. Even trusted access programs reduce the risk, but do not eliminate it: outputs leak out, techniques spread, and similar attack tactics will inevitably emerge elsewhere.»
For companies, this means that it is not enough to rely on the security precautions of AI providers. An independent security strategy based on current threat information remains essential.
Speed determines success or failure
Keizers paints a clear picture of the future security situation: «Defenders must now prepare for a threat situation characterized by widely available advanced AI capabilities. Success will go to those who simply act faster: Information must lead to a decision, and this must be followed by consistent action.»
The crucial question is not whether AI is used in defense - this is already happening - but whether companies retain control: «Those who do not actively shape the response are not automating strength, but uncertainty.»
Filigran is a cybersecurity company founded in France in 2022 and offers open-source AI-based solutions covering end-to-end threat intelligence management, attack simulation and security posture review. Over 6,000 organizations worldwide rely on the company's platforms.
Source: Filigree
Editor's update from 15.6.2026: By order of the US government, Anthropic has had to deactivate its two new AI models again. The reason given was that the security concerns were too great. Foreign nationals should not have access anyway.
This article originally appeared on m-q.ch - https://www.m-q.ch/de/wenn-neue-modelle-cyberkriminellen-in-die-haende-spielen/
