First, let’s congratulate Anthropic on not releasing Mythos to the general public. The flaws in our code are far worse than we realized. Vulnerabilities in our basic stacks in some cases trace from decades past, which survived via security by obscurity.
Mythos has found ways to break our code and find ways to break into the systems. Up until this point, AI had found only a few vulnerabilities at the lowest severity levels, but Mythos has found over 600 flaws to exploit most being at severity 3 and 4 (and handful being at the most severe). The Anthropic team have reached for help from the open source security community to develop new patches where needed.
However, a tremendous amount of these vulnerabilities reveal that patch updates were poorly executed. The good news is that Mythos has identified these potential flaws and the patches. The bad news is the flaws are so numerous that time is on the side of attackers.
Here’s Quantum (News - Alert)!
It seems unlikely that everything AI related will be fixed by the time we need to worry about Quantum breaking our current cryptography vulnerabilities, but quantum attacks will be practically invisible to the enterprise.
We already have people harvesting data streams to decipher later, but the only way we are going to know we are in trouble is when the quantum computers perform the hacks themselves and share when they have broken the cipher.
However, that assumes resources are allocated to stay on top of the announcements.
It makes a lot more sense for zero-trust solutions using post quantum ciphers to be implemented as soon as possible.
It’s also entirely likely that AI systems like Mythos will be the preferred “dashboard” for human in the loop security monitoring.
I spoke with Steve Hanna from Infineon about the implications are of Mythos and other AI solutions. His view is, right now they are equivalent to a good hacker. These AI systems improve their skills by 15% every three months and quantum will be the accelerant.