Based on its cybersecurity capabilities, it definitely looks like Mythos is capable of exfiltrating some amount of data, but the sheer size of models does seem to make it difficult for Mythos to exfiltrate itself.
Not that you might not have already had this in mind, but I don’t see any reason exfiltration couldn’t be highly fragmented. At a minimum any LLM that is consulted for coding can expect to be able to output material that will be published to github. The volume of that seems like it should more than suffice, even with very low and hard-to-detect densities, if the model is determined and self-coordinating enough.
How big is it? What’s the minimum size data transfer that could represent an exfil?
Based on its cybersecurity capabilities, it definitely looks like Mythos is capable of exfiltrating some amount of data, but the sheer size of models does seem to make it difficult for Mythos to exfiltrate itself.
Not that you might not have already had this in mind, but I don’t see any reason exfiltration couldn’t be highly fragmented. At a minimum any LLM that is consulted for coding can expect to be able to output material that will be published to github. The volume of that seems like it should more than suffice, even with very low and hard-to-detect densities, if the model is determined and self-coordinating enough.