You’ve heard of jailbreaking a phone. The same idea applies to AI models, and it’s a real business risk.
The idea
LLMs learn from massive internet data, including unsafe material, and rely on safety filters to keep it out of responses. Jailbreaking finds a way around those filters. For a business, that can mean brand-damaging outputs or even data exposure if the model is connected to internal systems.
One simple method, SequentialBreak, hides a harmful request among harmless ones. Because most of the prompt looks safe, the filter treats the whole thing as safe. Think of a 20-page contract with a clause on page 14 giving your supplier 10% of your equity. Alone, you’d reject it. Buried in routine language, it slips through.
Why it matters
- Filters aren’t foolproof. Clever framing can get a model to return what it normally blocks.
- Connected AI raises the stakes. Access to internal systems turns bad output into potential data exposure.
- Guardrails turn risk into an asset. With the right protections, you can build on AI confidently.
Spot the hidden clause
Like a good lawyer catching that buried clause, your AI systems need guardrails designed to catch what slips past the model’s own filters.
How are you thinking about AI guardrails in your business?
Join the conversation: see the comments on LinkedIn.
Read the full transcript
Hey, I’m Matt.
You may be familiar with jailbreaking phones, which is cracking the phone’s security so you can do things the manufacturer never intended you to do. Well, something similar can happen with AI models, and it’s called LLM jailbreaking.
LLMs are trained on massive amounts of internet data, which includes unsafe material. The LLMs rely on safety filters to stop that content from being returned. Jailbreaking LLMs is when someone finds a way around those filters, so the model produces answers it normally would refuse to return. For a business, that creates real risks — from harmful or offensive outputs that could damage your brand all the way to potential data exposure if the AI model is connected to internal systems.
One simple LLM jailbreak method is called SequentialBreak. It works by surrounding a harmful request with a bunch of harmless ones. Because most of the prompt looks safe, the model’s filter can be fooled into treating the whole thing as harmless — and that’s how it can end up returning dangerous material that the LLM would normally block.
Here’s an analogy, imagine you’re reviewing a 20-page contract. Everything looks normal, you see: payment terms, delivery schedules, and lots more typical legal language. But on page 14, hidden in a dense paragraph, there’s a line that says your supplier gets 10% of your company’s equity. If that line came by itself, you’d reject it immediately. But buried in routine language, it can slip past.
That’s exactly how SequentialBreak works with AI. Someone hides a malicious instruction inside a normal-looking request, and the model can get tricked into responding.
At SPR, we’re building systems to guard against these type of risks, the same way a good lawyer would spot that hidden clause in a contract. And when you are protected with the right guardrails, AI shifts from a risk to an asset you can confidently build upon.
What about you — how are you thinking about AI guardrails in your business?
Leave a Reply