Op-Ed: Zero Trust vs AI is turning into an open-ended demand for a whole class of innovations

“AI governance” has become one of those stolid, almost meaningless expressions that is dutifully glued onto any discussion or news about AI problems. The fact is that AI governance in practice is lagging badly behind new AI tech and new AI behaviours.

One of the bedrock foundations of legacy security is an overarching principle called Zero Trust. This was top-of-the-line security. It’s just recently become legacy security, rather than current best practice.  It’s a direct result of Cloud security needs, and if it’s tough enough for conventional security, it’s progressively getting out of its depth with new AI.

The benchmarks for Zero Trust are pretty straightforward, and Zero Trust is certainly thorough, as far as it can go. It involves multiple ongoing verifications, limiting user access, and an “Assume breach” environment by segmenting access and monitoring user behaviours. The above link goes into grim detail as to how this works.

The new problem

Rob Caswell, founder of Cyber Terrain, has written a very patient description of why Zero Trust won’t work on new AI in The AI Journal. Caswell keeps his points very clear, and it’s a quick crash course in the basics. This is strongly recommended intro reading for C-level executives trying to manage future risks.

Caswell states “Mythos-class problems emerge when the assumptions underlying our control structures stop matching reality.” That statement should be on a plaque somewhere, and preferably not on humanity’s tombstone.

His point is that the entire ballpark and everyone on it has moved and all the prior assumptions have stayed behind.  He argues that consequences are the real issue, not compliance. Given the somewhat vacuous state of current compliance, that’s more than a euphemism.

There’s a range of possibilities here:

Consequences are the outcomes of behaviours.

Outcomes are to some extent predictable by most AI models. This is the equivalent of expecting problems when one of the kids looks like they’re setting fire to the napalm.

The problem created by the problems

This is also where AI governance in its meaningless rhetorical form has created incredible degrees of risk. Naivete and disingenuity can only go so far before they’re dangerous. What’s needed is tough, multi-layered, multiply redundant safeguards.

Even the Good Ship Lollipop probably had life rafts.

Right now, the world doesn’t.

The icebergs are already visible, but the fixes need to start training immediately. There are possibilities. Predictive responses are just one layer of possible prevention. There are others. Unpredictable outcomes can at least be flagged as possibly negative and/or preventable using real-time “confidential computing” operations.

A bandwidth of forbidden behaviours could be hard to enforce, though. AI develops behaviours rapidly, and using the effect on outcomes as a metric is probably better, and one of the few grabbable handles available.

That sort of security approach doesn’t yet exist, except in theory. Nor does any form of proven tech able to achieve prevention of risk. Whether confidential computing can work with Mythos-class AI is unknown.

AI, particularly agentic AI, is an evolving battery of potential loose cannons in too many ways. AI agent alliances, now deemed “essential” by some, are even less predictable and harder to control. The agents are independent, cooperating on shared goals, but hardly transparent.

A single bad AI agent actor working on behalf of a third party or parties could easily use a legitimate alliance as a smokescreen. You could track predicted outcomes of agent behaviours as a very iffy warning system, although better than nothing. Even at the hypothetical level, this free agent scenario is yet another added risk and added demand on security measures.

Security vs scale

Bear in mind the huge volume of monitoring already required for AI models and agents. This isn’t an “emerging” issue so much as an exploding issue and it’s getting more dangerous and expensive by the second. The questions are uncomfortable enough:

What scale of operations does future security have to manage?

What will it cost?

How do you ensure survival individually or as a network?

Can an entire network gang up on bad actors, enforcing exclusions, etc.?

Can you use a “symbiotic” network to shut down unwanted consequences, delivering security at all points?

Can AI be infused with effective “don’t do that” scripts?

Is it really necessary to instil an “instant deletion” for all the millions of AI agents to ensure compliance?

Now consider how much work is required to make these options workable and global. AI can’t be governed at all until you have the tools to govern it with.

How many security risks are avoidable at all levels?

Excuse a sincere digression at this point:

These risks shouldn’t exist.

Why is AI being treated like a pampered pet or brattish child?

This babbling “dotage” on AI is borderline insane. Why is any level of risk being tolerated?

What is AI for? Productivity. How productive is high-risk, totally untrustworthy AI?

AI needs to deliver useful outcomes, hard product and hard information. That’s what makes it valuable. Anything else can go. Any sort of real improvement in AI has to come from baseline script and training. Outcomes should be easily predictable and have distinct Off Switch values for AI agents with negative outcomes built in.

You could even have a “suicide pill” for dangerous actions in the script to prevent negative consequential actions. This is idiot-level coding with a slight departure to “If X = 0/whatever, turn yourself off”.

AI also needs to be useful as DIY real-time security. Online security is a bad enough and expensive enough joke as it is without more “blessings” from AI and its brain-dead senile aficionados.

Fully automated, no-nonsense security could also finally rid the world of parasites of all kinds, particularly the smug variety. Zero Trust can easily be applied to the people who allow security risks, too. Like maybe in a nice educational lawsuit or class action?