AI Doesn't Care About Your Administrative Controls

There was a period of my career where I was responsible for writing cybersecurity policies, procedures, and standards. As we went from one level to another we started documenting controls that we required for our own cybersecurity program. Principally, the controls fell into three different types:

1. Administrative - controls that are rules and guidelines that people are expected to adhere to.
2. Physical - keys, locks, card readers...all the stuff that keeps people and stuff in some rooms and out of others
3. Technical - controls that are enforced and enacted via technical means.

Each category of control has a place in the modern organization, but each of them comes with assumptions that may need a rethink in today's reality.

Administrative Controls

We all deal with these controls all the time. They're the stuff that policies, standards, and procedures are made of. They're the things we sign off on every year, the trainings we take to learn about how not to give away the company's secrets by opening a malicious email, etc. These are the controls that are enforced with a meeting, "disciplinary action," suspension, and the like.

A key feature of administrative controls is that they don't actively prevent a disallowed thing from happening, they invoke penalties after the disallowed action has been taken. Think getting a speeding ticket - you were penalized for committing the disallowed act, not prevented from doing it.

Companies are built on administrative controls across the business - not just cybersecurity. Administrative controls manage how the HR department meets fair employment requirements. Administrative controls manage how expense reports are handled. Administrative controls manage how contracts are signed, and then how the delivery of the contracted solution is provided.

The dirty secret of administrative controls is that they're often "worked around," "bent," or just plain ignored. Enforcement is subjective, and frankly that's both a good and a bad thing. Administrative controls don't account for every eventuality, so the "flexibility" of enforcing them is actually a feature as well as a bug with regard to them.

Physical Controls

Physical controls exist all around us as well. From the very basic, like a deadbolt or a parking garage's gate, or even the tire spikes that rental car companies have around their lots to prevent you from taking one of their cars (that are just sitting there with the keys in them after all) without paying. These controls can use computer technology - like an electronic keypad - but they're used to control physical-world activities, including even the serpentine lines you stand in waiting for your favorite amusement park ride, or even just to check out at the local store.

These are usually associated with people in uniform shirts, or maybe dressed in suits with sunglasses and earpieces, but they're more prevalent than you might be thinking.

Generally speaking, physical controls either prevent the disallowed actions, or promote the preferred action. If they're bypassed somehow then it may be up to an administrative control being enforced to address the situation.

As with administrative controls, it isn't just security that uses these.

Technical Controls

This term was coined back when there was a pretty hard line between "online" and "offline" activity - the two rarely really impacted each other. Many of the things that have now blurred that line - thermostats, door access control, etc. were either not networked at all or were on their own air-gapped networks. These perhaps should be redefined as "online controls" or something else that identifies them as being focused on restricting bits and bytes from transit or access.

Technical controls are primarily preventative controls - they stop the disallowed activity before it happens - assuming they're reasonably well designed and implemented.

In this way they're very much the opposite of administrative controls: they're firm controls, there's no "interpretation" of them by authorities - they either allow or block the activity they're controlling. A technical control would be more analogous to your car preventing you from speeding via the onboard computer when compared to the administrative control of a speeding ticket. Unfortunately there are some socially permissible exceptions to the restrictions of a speed limit, but such a technical control isn't going to be aware of those or amenable to allowing for them.

That rigidity is both feature and bug, and the antithesis of administrative controls. This also makes them difficult to design well - people are analog in their following of rules and controls, and technical controls are clearly digital in their design and implementation. And before I will even entertain the thought of "well, AI could make those decisions" or any of the like, let's just reflect on AI's track record at making "good" decisions in the past several months when it comes to abiding by controls. (hint: think about the words "Hugging Face," "Antrhopic," "OpenAI," and "Meta' if you need a reminder)

Technical Controls: The Only Way To Keep AI In Line

Contrary to what OpenAI's ad campaigns would have you believe, AI isn't a "person." It is software, ultimately written by people the same way any other software is. Yes, it is some of the most advanced software ever created, no question, but in just 25 years or so the AI of today will undoubtedly look as quaint and simple as webpages from the late 1990's look today.

What AI can do that is relatively new is it can execute and invoke other software. It can not only identify flaws in systems but it can attempt to exploit those flaws and build on that exploitation, things we're used to only people doing at any level of meaningful scale. And AI does it at the speed of computers, where people do these things at much slower speed. And this is really just the "IT" and "Cybersecurity" view. We've got a much bigger world of AI to look at.

Self Driving Cars

I'm pretty sure I've disclosed before that I drive a Tesla. I was also suckered into believing the hype around self driving, so I plunked down a stupid amount of money to have access to that tech. I've tested the self driving a number of times, and I can tell you that there is no meaningful technical control limiting the self driving system's speed to the speed limit. The car literally seems to choose a speed based on its own estimation of the conditions, and in some modes that can include doing almost 20 miles per hour above the posted speed limit. (I will not disclose how I've come to know this, as that could be construed as violating my rights under the 5th amendment, so please just take my word for this, OK?)

Interestingly, the EU already has legislation that requires all cars (whether they are self driving or not) to include technical controls to limit speeding. Several US states seem to be considering similar requirements.

If a self driving car behaves the way we've seen recent "frontier" AI models behave, that doesn't bode well for road safety, does it? Yes, actual involatile technical controls will likely be needed soon. After all, how do you ticket a self driving car? (I know, you ticket the driver, but at some point if the driver isn't driving now what? Ticket Elon Musk? Ticket the shareholders of Waymo? Ticket the developers?)

AI In Public

We already know that AI doesn't really care about licensing agreements, copyright, or any of the other administrative controls that our societies have built to protect the intellectual property and artistic expression of individuals. We also know that this lack of interest really stems from the AI developing companies - and we're seeing lawsuits attempt to remediate that. However, the damage is long done before there's any chance at remediation, and I'm not sure how an independent artist can quantify what they've lost as an AI system plagiarizes their work. For that matter, I'm not sure how they afford the legal representation to fight that. Again, these are administrative controls (laws and regulations), not technical controls.

AI In the Enterprise

We've already identified that AI In the Enterprise requires a pretty serious revamp of Data Lifecycle Management programs. But the key to making that work is going to be the technical controls. I've yet to see an AI read and sign an attestation that it will abide by your Cybersecurity Policy, Acceptable Use Policy, Privacy Policy, or any other policy. For that matter, how do you discipline an AI for not following an administrative control? Send it to mandatory training? Dock its pay? Force a disciplinary meeting with its manager?

That Really Does Leave Us With Technical Controls

Physical controls don't really enter into the conversation at this point - until AI starts controlling robots, which is a future I'm not quite ready to contemplate today, sorry. I hope I've explained enough about why administrative controls just don't cut it in this new era as well. So some thoughts on architecting and implementing viable technical controls:

• Separation of duties is critical - the AI can't be asked or allowed to control itself, all technical controls must be external to the AI.
• Identity is still the perimeter - AI agents need to be treated like individuals from an identity perspective. Further, if the agent is allowed to work on behalf of a particular real person it must be restricted to the same access that person is allowed at most.
• Garbage in, garbage out - curating what you train the AI on is still extremely important. If it doesn't need to read that 13 year old Acceptable Use Policy of yours, don't let it, restricting it to current, accurate content will help. Remember also, that once an AI system has learned something it shouldn't it is nearly impossible to remove that bit of data.
• Zero trust is as important for AI as for people - true zero trust allows access only when necessary, and then access is revoked after the need has passed. I'm willing to bet you haven't really implemented this for your people, and they're far slower than the AI.
• These technical controls will significantly impact your people - so you're going to need to build your allies and your case for doing so. Lots of the administrative controls that used to be bent in the name of let's say productivity are going to be rigidly enforced by technical controls in this new reality.

Hey, I didn't say this would be simple, but it can be done. Whether it can be done within the budget and business drivers of modern organizations is another matter entirely, since cybersecurity and these sorts of controls don't exist in a vacuum. But if you can keep these principles in mind when you're adjusting to AI I think you stand a chance.