Claude Mythos #2: Cybersecurity and Project Glasswing(thezvi.substack.com)
thezvi.substack.com
Claude Mythos #2: Cybersecurity and Project Glasswing
https://thezvi.substack.com/p/claude-mythos-2-cybersecurity-and
https://thezvi.substack.com/p/claude-mythos-2-cybersecurity-and
According to the cybergym benchmark in the article, Mythos is 83% of the way towards being perfect, not 100%, and there will still be holes in critical code after Mythos has done it's magic. Also, if Mythos can find exploits so can competing Cutting-edge AI models, some with nefarious actors.
A strategy for the next 12-24 months at least would be to assume ALL your systems will be compromised, move critical data to air-gapped backups, and never use the same system for browsing the web and for banking or other critical stuff.
I predicted similar scenarios in a sci-fi book I wrote a few years back, though we may have skipped the part where Mythos would be spending its first GPU cycles improving itself first :-)