Once more, I am writing about AI. By now, I have become a nearly daily user of Claude Code. I have learned how to work with it. I learned to structure the work so that my LLM agent does not run into the Context Rot problem. Maintaining the CLAUDE.md file. Every time it makes a mistake, the question is: what do I need to teach it in my MD file structure so that it knows better the next time? So, in a way, I have become an Anthropic fan.
But recently, I was reading about Anthropic’s newest model, which they call Mythos preview. My first thought was that it was perfect marketing. Anthropic wants to create a myth around this model, so they call it Mythos. Makes sense. But it is also a bit over the top.
The story is that it is highly effective at finding software vulnerabilities and creating exploits for them, which makes it dangerous if bad actors were to get hold of it.
Therefore, Anthropic decided that the Mythos preview model is not available to regular customers or the public. It is available only to specific hand-selected organizations that are part of ‘Project Glasswing.’ On the list are several big names, including Apple, AWS, Cisco, Google, Microsoft, and The Linux Foundation.
I see the point. It makes sense to me. I appreciate that Anthropic has higher ethical standards than some other AI players.
However, the issues are in the details. How is/will the Linux Foundation filter which maintainers get access to the dangerously effective Mythos model? Does Anthropic bind its partner companies to any standards on how they select which employees get access to it?
Will job seekers in the future ask their potential employers for details when they get access to such models with restricted availability?
Will it be a crucial skill for a future cyber criminal to work for a high-profile company in their day job and keep some of the exploits they find for their own criminal use?
These concerns make the ‘insider threat’ even more dangerous.
Moving over to the latest LINBIT content, ‘Integrating External LINSTOR Clusters With Kubernetes by Using the LINSTOR CSI Driver’ is a blog post that helps people deploy LINSTOR outside of Kubernetes. It includes a lot of useful information and links.
There has also been another collaboration between 45Drives and LINBIT since I last wrote. This time, ‘100% Uptime for Less: Building a 2-Node DRBD Highly Available Cluster with Pacemaker/Corosync’ is a demo about ‘building a high-availability (HA) cluster using open-source tools that won’t break your budget.’
Finally, since I last wrote, we’ve released ‘drbd-9.3.2 and drbd-9.2.18.’
