CogSec

With the advent of models with high cyber capabilities like Mythos, the tech sphere has been increasingly concerned with CyberSec and OpSec.

One other threat vector that such models greatly amplify, which is perhaps more insidious, is towards CogSec—cognitive security.

Information hazards, or infohazards, have been discussed at length in sci-fi. However, they increasingly exist on the real-life internet. In sci-fi, this is mere information, e.g. memes, images, text, etc. which affect the consumer in ways ranging from amnesia to instant brain death. In real life, infohazards manifest as low-entropy slop disguised as high-entropy information worth consuming. Such slop is emitted either by LLMs, or proids¹.

Examples:

• Every LinkedIn post
• 90% of posts on AI Twitter
• YouTube comments
• Probably a lot of YouTube videos
• Fox News
• (Most) short-form video
• Many LLM outputs, even from frontier models, even if they are factually correct, due to slop writing and slop framing. Also, even when models aren’t slop, slop users direct them to produce slop which they plaster over the internet
• etc.

Infohazards gunk up valuable information processing channels. I postulate the following:

Infohazards of this sort have probably existed since the printing press, but the dual proliferation of the internet and LLMs means infohazards probably now outnumber infogems² in the average person’s information channels.

For the sake of your CogSec, you must shut out as many infohazards as possible. I truly believe that unless the sci-fi scenarios of AI doom or utopia play out in the next decade, the most important thing one can do is prevent their brain from melting into soup under a torrent of infohazardous slop. Those retaining higher-order brain functionality will form the true overclass.

AGI is here, and you still don’t have a CogSec threat model?