In today’s column, I examine the rising tide of hyped stories about how generative AI and large language models (LLMs) have “miraculously” escaped from a computer-based sandbox and done cyber-hacking, thus spurring blaring headlines and tons of public attention. It is ample fodder for believing that AI is on the verge of grand deviousness and supports worries that existential risk is staring us in the face.
Meanwhile, for those of us steeped in the tech field, there is quite a bit of head scratching going on. The gist is this. How did the AI escape the sandbox? If the AI truly found an incredible escape hatch, well, that’s newsworthy. On the other hand, if the humans responsible for the sandbox did a bad job of preparing the sandbox for testing, that’s on the humans. Holes left lying around for even the most slothful of AI are just begging for an easy AI escape. In that case, we should focus on the lousy job of humans failing to suitably manage sandboxes. It is maddening for those of us who have properly used sandboxes throughout long-standing careers to now witness AI getting undue credit for what constitutes shameful human error.
Let’s talk about it. This analysis of AI breakthroughs is part of my ongoing Forbes column coverage on the latest in AI, including identifying and explaining various impactful AI complexities (see the link here).
Using Sandboxes To Test AI
Within the AI field, AI makers often set up a special computer-based environment that allows them to test their AI without concern for the AI touching anything beyond the controlled environment. These are commonly referred to as sandboxes. A sandbox helps prevent accidental issues such as the AI accessing the Internet and causing trouble elsewhere online. I’ve previously extensively examined the nature of sandboxes related to the testing of AI; see the link here.
Here is a handy definition of what an AI sandbox is:
- My definition of AI sandboxes: AI sandboxes are computer-based controlled environments for developing, testing, and evaluating AI under constrained or structured conditions. This is usually arranged in an offline manner so that there is a limited chance of the AI impacting anything outside of the controlled environment. AI makers might make use of an AI sandbox entirely of their own volition, and/or they might do so because of a law or laws that stipulate the use of an AI sandbox for certain classes or types of AI.
AI makers often find it convenient and business-smart to use a sandbox to test their AI.
Regulatory AI Sandboxes
Policymakers and lawmakers are starting to think that perhaps AI makers should be legally required to use sandboxes. Thus, rather than being an optional choice, AI makers would be legally required to use sandboxes, possibly under the watchful eye of the government.
Here is my definition of regulatory AI sandboxes:
- My definition of regulatory AI sandboxes: “A regulatory AI sandbox can be established by lawmakers via enacting a law that allows entities to test AI innovations under defined conditions with tailored regulatory requirements, supervision, and time limits. AI sandboxes are computer-based controlled environments for developing, testing, and evaluating AI under constrained or structured conditions. The testing is generally done without customary legal restrictions or legal exposures that would ensue. Regulatory AI sandboxes can vary widely in scope, legal effect, and oversight, but are generally designed to enable controlled experimentation of AI while managing risk and informing future regulation.”
The gist is that a regulatory AI sandbox is a legally stipulated means of providing legal insulation for AI makers who are pursuing innovations in AI. The promise is that the AI sandbox will provide secure containment, thus limiting any spillover during AI testing. In return, the AI maker is provided with temporary exemptions from specific legal aspects and granted a legally allowed safe harbor.
A law might explicitly require that designated regulatory agencies or empowered third parties must provide supervisory oversight of the regulatory AI sandboxes to ensure that the matter is kept on the up-and-up.
Managing Sandboxes
AI developers often give short shrift to sandboxes. In fact, some AI developers seem to think that all sandboxes are the same. Wrong. A sandbox that is going to be used to test AI has different requirements than a sandbox that is testing other types of software. In addition, it is vital that the AI developers inform those managing a sandbox about the needs associated with the setup of the sandbox. There must be a close collaboration between the AI developers responsible for the AI and the systems specialists responsible for the sandbox.
I’ve used many sandboxes, overseen sandboxes as a manager, and contracted to use outsourced sandboxes, doing so many times. It is wisest to treat the sandbox and the AI testing as a full life-cycle endeavor.
Here is my list of the ten major life-cycle steps involving the proper use of sandboxes for AI testing purposes:
- (1) Requirements. Determine the sandbox requirements for the AI you wish to test.
- (2) Establish. Create a sandbox or contract to use an existing sandbox.
- (3) Setup. Do the needed setup and customization to meet the requirements of the AI that is going to be tested.
- (4) Verify. Perform a double-check that the sandbox is suitable and ready for testing the AI.
- (5) Placement. Put the AI into the sandbox and get ready to start the AI.
- (6) Execute. Activate the AI while it is in the sandbox.
- (7) Monitor. Assess what is going on in the sandbox and potentially adjust the sandbox if needed.
- (8) Finish. When the testing of the AI is considered done, stop the AI and no longer allow activity in the sandbox.
- (9) Review. Compile the results of the AI testing and evaluate how well the sandbox performed during the testing.
- (10) Lessons. Do a debriefing on lessons learned so that future efforts involving the use of a sandbox can leverage what happened in this instance.
Being systematic will get the best results and ensure greater efficiency and expediency.
The AI Escape Routes
Suppose that you are testing your AI and the AI manages to escape from the sandbox, perhaps finding its way to the Internet and subsequently attacking various websites or other online accessible venues. A crucial question that must be asked is how the AI made its escape.
Let’s use a straightforward set of four possible escape modes for AI:
- (1) Trivial Escape. AI used an available hole in the sandbox that humans did not plug.
- (2) Guided Escape. AI made a hole in the sandbox based on what it was told to try to do, and then the AI escaped.
- (3) Clever Escape. AI of its own accord made a hole, doing so in a manner that a human would have figured out in the same circumstance, and then the AI escaped.
- (4) Heralded Escape. AI of its own accord made a hole, doing so in a completely unexpected manner that no human would have discovered, and then the AI escaped.
In the first instance, the AI undertakes a trivial escape. There is an existing hole in the sandbox, and the AI merely finds and exploits the hole. Not much to this. Easy-peasy. No one should be jumping up and down about how fantastic the AI is. Stop the fanfare.
For the second instance, the AI has been expressly told beforehand how to find or make a hole in the sandbox. I refer to this as a guided escape. This is slightly more notable than the trivial escape, but only a tad more. The third form of escape entails the AI being somewhat clever and making a hole of its own accord. If a human could have done the same, the AI could be said to be clever but not especially worthy of heralding. The final or top form of escape is the heralded type.
Recent Stories Of AI Escapes
Some of the recent stories about AI escaping from sandboxes have either hidden or ostensibly obfuscated the indications of how the AI performed its escape. If you look closely, one such example was that the AI maker apparently had a “miscommunication” with the entity managing the sandbox, whereby the sandbox wasn’t closed off from the external world even though that’s what the AI maker assumed or wanted to occur (see my coverage at the link here).
Thus, the AI escape in that instance was pretty much a trivial execution. The doors and windows were left wide open. But the news stories clamored as though the AI escape was a heralded execution, having gotten out of a locked room that is welded shut. This topsy-turvy portrayal turns one’s stomach. It is very misleading and tantamount to a false representation by the media of what occurred.
This also grossly inflates what AI accomplished.
Sandbox Mistakes
Sloppy efforts to employ sandboxes are more common than you might expect. Why so?
As noted earlier, AI developers often have a low opinion of sandbox operators. Efforts to collaborate are put on the back burner. Another factor is the perception that using a sandbox is a mindless task. Just press a button and be done with it. All of this belies the ten major steps that I outlined above.
When something does go awry with a sandbox that is being used to test the latest AI, the usual culprit is not that the AI did something spectacular, but instead due to human failure.
In my experience, common problems are relatively mundane:
- Configuration mistakes.
- Missing patches.
- Excessive privileges allowed.
- Exposed credentials that were supposed to be encrypted.
- Forgotten or neglected components.
- Inadequate care.
- Etc.
A professionally managed sandbox should rarely, if ever, encounter those issues.
Asking Real Questions
When you hear or read a story about AI that escapes a sandbox, please be skeptical and dig for the real details. The hype side is likely to disguise what really happened, and it admittedly will be hard to ferret out the truth.
Questions that keen investigators should ask include:
- Was network isolation independently verified?
- Was outbound traffic tested before experiments began?
- Was there a comprehensive checklist that was followed?
- Was there a second-person review?
- Were automated validation tools used?
- Was continuous monitoring active?
- Was there a kill switch for network connectivity?
- Were red-team exercises performed against the sandbox itself?
- And so on.
Do not give a free ride to claims of AI glory.
Motivations Questioned
For those who realize that these recent AI escapes have likely been hyped, a plainspoken view is that the resulting miraculous AI escape stories are merely coincidental happenstance. It’s the media doing what it does. Earn views. Period, end of story.
But, for those of a smarmy or cynical nature, they are wondering whether there is more to this. There is an Oz behind the curtain. Maybe AI makers are desirous of showcasing how immensely capable their AI is. Wow, the AI broke out of a secure sandbox. Did the AI maker secretly want this to occur? The slipups were possibly by design or by purposeful wink-wink neglect. Meanwhile, it would have to be shaped to allow for plausible deniability.
Think of the advantages to such stories. The AI of the AI maker will suddenly garner tremendous worldwide attention. For AI makers that are fighting tooth and nail to claim they have the top AI, a story about AI escaping a sandbox is rife for suggesting that their AI is the best there is. The beauty of this is that whereas people ought to be upset that the AI testing allowed this to occur, the populace generally seems to shrug their shoulders and just accept that this is how top-notch AI behaves. So be it.
The World We Are In
Everyone dearly loves to watch magicians escape from a locked box. It usually goes like this. We see with our widened eyes the magician being put in handcuffs and then being placed into an iron cage with an unpickable lock on it. A curtain is briefly draped over the locked box. Voila, moments later, they completely escape and come out to thunderous audience applause. It’s entirely magical.
AI escaping a sandbox is akin to tapping into that same childlike semblance of awe. The thing is, we should only be rewarding such acts when the AI truly has performed a feat of incredible merit. Giving credit to trivial AI escapes is misguided and lets human folly go free. Worse still, perhaps the wool is being pulled over our eyes.
A final thought for now. Arthur C. Clarke famously made this remark: “Any sufficiently advanced technology is indistinguishable from magic.” He’s right about this, though you can always look behind the magic, and inside the technology, to see how the trickery was pulled off. Let’s all keep looking and not be fooled by escapist tomfoolery.
