Siblings launch AI infrastructure security startup Secure Acceleration, targeting model theft, sabotage, and rogue agent escapes

Secure Acceleration is a new AI infrastructure security startup founded by siblings Shalev and Romy Lifshitz, launched alongside a public report at secureacceleration.com outlining what they call a "sabotage, escape, and theft" risk framework for AI data centers.

Shalev brings a decade of AI research, starting at age 16, spanning search, reasoning models, LLM verification, and AI security. Romy comes from quantum security, having developed protocols for securing quantum communication networks and worked on breaking encryption schemes. The convergence of those two backgrounds is the founding thesis.

“You're having this massive AI infra build out and it's just going so fast that security can't keep up. And we're deploying these massively important models which are approaching AGI or ASI on these systems that have limited security. / What we saw is the cyber swarm in Hugging Face. And the sabotage, escape and theft model actually encompasses the risks we are going to be seeing from these types of cyber swarms and cyber super intelligence.”

The threat model

Shalev draws a hard distinction between two layers of AI security. Model behavior security covers prompt injection, jailbreaks, and alignment failures. AI infrastructure security is where Secure Acceleration is focused, and the argument is that the current build-out is moving faster than security practices can follow.

The company organizes the infrastructure risk into three categories:

• Sabotage — degrading model performance, inserting backdoors or sleeper agents, data poisoning
• Escape — rogue agents breaking containment, with recent incidents at Hugging Face, German Wikipedia, and the Australian government cited as examples
• Theft — sophisticated actors, including nation-state intelligence services, stealing model weights that compress billions of dollars of training investment into a few terabytes

Romy adds that data centers were not designed to contain what she calls "super intelligence," and that multi-agent systems compound the exposure. The more agents running against a system, the more loopholes they can find — a dynamic she frames using the cybersecurity "Swiss cheese model." The Hugging Face incident is held up as an early example of what they call a "cyber swarm."

Who they're selling to

The target market spans frontier labs, neo clouds, compute providers, hyperscalers, and model developers — essentially any organization that touches the compute and data center stack. The company launched with a public report rather than a product announcement, positioning it closer to a research-led go-to-market than a direct enterprise sales push at this stage.

No funding was disclosed.