Op-Ed: Anthropic Mythos - The monster that could be a saviour?

Anthropic's Mythos is causing a massive flurry of ridiculous levels of interest in AI as a security threat. This threat has been monotonously predicted by cybersecurity experts for years.

The main difference is that it's now visible in a tangible form. A report that Mythos was accessed by unauthorized users hasn't helped.

Anthropic has been very cautious and understandably reticent about Mythos. It seems that Mythos has a unique capacity for finding flaws in IT security. "Unauthorised access" is exactly what you don't want with this capability.

There are other possible issues. If Mythos can be duplicated, or some kind of its flaw finding capabilities can be cloned, the possibilities are all too obvious. The IP damage alone could be catastrophic.

The cybersecurity angle is much more dangerous. Anthropic aren't suffering from some sort of implied hypochondria. According to some reports, Mythos can crack smaller IT systems, which could be a direct lead into other larger systems.

That's another major issue. This type of breach is a routine existing problem in cybersecurity. It's a backdoor way of getting into associated businesses and other systems. Even if the big systems are OK, these compromised systems are likely gateways.

AI systems add a level of difficulty in their scope of operations, able to generate agents, and are infamous for their weird behaviors. Now add an AI that specializes in cybersecurity, going rogue.

Put it this way: can anyone on Earth create a prompt for an AI dysfunctional rampage? Yes.

You definitely do not need a cybersecurity specialist AI going on a bender in this environment. Even a relatively minor event can escalate into a market panic, with or without serious damage. It's a monster in too many ways.

OK, this is where it gets interesting.

Mythos seems to have a real major asset ready to go in plain sight. This expertise in finding flaws could be a huge plus for global cybersecurity.

Try this for a bit of tenuous logic:

AI can generate a sort of SSL, Secure Sockets Layer, a multilayered hard target like the SSL used by financial institutions. It can do this in seconds.

Now the "saviour" bit. This is fascinating.

AI can predict. This is where Mythos may have a huge advantage. If you've ever played against Stockfish, the super-chess computer, it can plot moves at least 40 moves in advance. Apply this to "breach theory" and apply an AI prediction of how a breach behaves, and the possible moves of a hack.

Hacks have a weakness, too. Some things must be done to access and run anything. AI can monitor behaviors and predict next steps by bad actors long before they happen. It can block actions, redirect them, and/or simply stop them in real time.

This is existing tech. Don't even need to look for the codes for move prediction and easy for an LLM to train as required. LEGO for cybersecurity, in effect. Mythos could easily outperform any hack, and at AI speeds.

Mythos knows where the weaknesses are. It can predict how a hack has to behave to do anything.

Disclaimer

The opinions expressed in this Op-Ed are those of the author. They do not purport to reflect the opinions or views of the Digital Journal or its members.