Everyone Had an Opinion on AI Security. Almost Nobody Claimed Control.
Last week in New York, Hubscale and Evolution Equity Partners brought together a room of cybersecurity and AI leaders for Off The Record.
The panel focused on one of the biggest questions facing security teams right now: as AI adoption accelerates, are businesses actually in control of how it is being used?
The discussion covered everything from whether AI security should have its own budget line, to where dedicated AI security platforms are genuinely delivering value, how attackers are using AI, where model guardrails fall short and who is ultimately accountable when things go wrong.
There were different views across the panel, which made for a pretty lively conversation.
Are businesses actually in control of AI?
It opened with a simple question: who actually knows what AI is being used inside their company, and who feels like they have meaningful control over it?
By the end, one hand was still up.
He worked for an AI security vendor.
That pretty much summed up the conversation that followed.
The panel brought together Pieter Vanlperen from AlphaSense, Chris Hughes from Noma Security, Kyle Novellano the CRO of ZeroFox and Jake Bernardes at Gambit Security, with Samantha Feuer from Evolution Equity Partners hosting.
There was plenty of disagreement, particularly around the idea of AI security becoming its own category.
There are clearly platforms solving real problems, but one of the points that came up was the difference between solving a specific security issue and claiming you can secure enterprise AI as a whole.
That gap still feels pretty big.
The problems aren’t all new
A lot of the problems being discussed weren’t exactly new either.
Permissions, exposed data, weak separation and systems that nobody fully understands have been causing headaches for security teams for years. AI just adds another layer of complexity and, in some cases, gives those problems more room to cause damage.
AI is lowering the barrier to attack
The attacker side of the conversation was probably one of the more interesting parts.
Jake described it as a potential “new age of the script kiddy”.
As AI gets better at handling more of the research and execution behind an attack, the level of expertise needed to get started could drop quite quickly.
Someone doesn’t necessarily need to know how every part of an attack works. They increasingly just need to know what they want to achieve.
That means more people can try, more targets can be tested, and the cost of getting it wrong and trying again keeps falling.
Who is accountable when models behave unpredictably?
The question of accountability also came up.
Pieter shared an example of an enterprise model randomly inserting spaghetti recipes into policies and procedures. Funny at first, but less so when you think about how much responsibility businesses are starting to give these systems.
According to Pieter, the provider couldn’t really explain why it was happening. The suggested fix was basically to tell the model not to do it.
That feels like a fairly weak answer when companies are trying to connect AI into important workflows and make it reliable enough to use at scale.
Guardrails only go so far
There was also some scepticism around native guardrails.
Attackers don’t have to use the newest model, the safest model or the one with the most restrictions. They can use whatever gives them the easiest route to the result they want.
That makes the challenge pretty simple on paper and much harder in practice.
More questions than answers
For all the different views in the room, the main takeaway was probably how much is still unclear.
AI security is moving quickly. Companies are buying tools, rolling out new systems and making decisions now, even though there are still big gaps in understanding around what works, where the real risks sit and who is ultimately responsible when something goes wrong.
There probably isn’t a neat answer yet.
But buyers can still ask better questions.
What problem does this actually solve?
What can it not do?
What evidence is there behind the claims being made?
And where does confidence turn into assumption?
We started the night with one hand in the air.
After the discussion, there were still a lot more questions than answers.
Which, for this topic, was probably the point.
If you found this discussion interesting and want to continue the conversation around AI, cybersecurity or the market more broadly, get in touch with us.