A New Security Startup Says AI Chips Have a Blind Spot. It Wants to Fix That.
Stealthium is building tools to spot attacks hiding inside the specialised computer chips that power artificial intelligence, a corner of corporate IT that most security software cannot see.
Key points
- Stealthium is a new cybersecurity startup focused on detecting attacks inside AI accelerators, the specialised chips used to run artificial intelligence workloads.
- Current security tools largely cannot monitor activity happening inside these chips, leaving a gap that criminals could exploit.
- The company targets "neo-clouds", smaller cloud computing providers built specifically to rent out AI processing power.
- Stealthium's approach analyses subtle behavioural signals from hardware to flag unusual activity that conventional tools miss.
Most companies running artificial intelligence have no idea what is happening inside the chips that power it. That's the gap Stealthium, a newly launched security startup, is trying to close.
The chips in question are called AI accelerators: specialised processors designed to handle the heavy maths AI models require. Graphics processing units, or GPUs, are the best-known type. They sit at the heart of almost every major AI service today, from chatbots to image generators, processing enormous amounts of sensitive data. Mainstream security software has focused on servers and software, leaving what happens inside these chips largely invisible.
Stealthium's pitch is that this blind spot is real and exploitable. The company says it watches for subtle telemetry signals, tiny streams of performance and behaviour data that the chip emits while working, and uses those signals to flag patterns that don't fit normal operation. Think of it like listening to a car engine for sounds that suggest something's wrong, rather than waiting for the warning light.
Who is most at risk?
The companies with the most exposure are what Stealthium calls "neo-clouds": smaller cloud providers whose entire business is renting AI chip power to customers. Traditional cloud giants have large security teams and years of hardened infrastructure. Neo-clouds are newer, often running hardware at scale before security processes have caught up.
Any business renting AI computing time from one of these providers, or running its own GPU cluster, sits in the same exposed position. If an attacker found a way to abuse activity inside accelerator hardware, conventional tools would likely miss it entirely.
No specific breach or CVE, meaning a numbered record of a publicly known software flaw, has been attributed to this class of weakness publicly. Stealthium's argument is that the absence of known incidents doesn't mean the risk is absent. It may simply mean nobody has been watching. Our 29 July story "Your Security Team Is Flying Blind on AI" made a similar point about AI agents operating outside the reach of conventional defences.
SecurityWeek first covered the company's launch and its focus on this hardware-level monitoring gap.
What should ordinary businesses do right now?
If your organisation rents AI computing power from any provider, ask that provider directly what security monitoring they have in place at the hardware level. Most won't have a satisfying answer yet, and that's useful information.
For teams running their own GPU infrastructure, check whether your security monitoring reaches the hardware layer or stops at the operating system. Knowing where your visibility ends is the first practical step. The honest read on Stealthium is this: the threat they're describing is plausible and underexamined, but without a documented incident to point to, the company is selling urgency on borrowed time.
