We have spent the last decade building smarter models. Now, the industry is pivoting toward making those models do things. The jump from a chatbot to an autonomous agent is a massive leap in utility, but it is an even bigger leap in risk. When an LLM hallucinates, it gives you a bad recipe. When an agent hallucinates, it drains your corporate credit card or deletes a production database.
This is the problem a new startup called Disseqt is trying to solve. Backed by former insiders from OpenAI and Meta, the company is positioning itself as the safety layer for the coming wave of autonomous systems. It is a classic infrastructure play: as everyone else rushes to build the engines, they are selling the brakes. And if you are building in AI right now, you need to be paying attention to how these brakes are being designed.
The Shift from Chat to Agency
For most builders, the novelty of basic prompt engineering has worn off. We are moving into the era of Tool Use and Function Calling. We want AI that can log into our CRM, manage our calendars, and interact with third-party APIs. This creates a security nightmare. Traditional software security relies on predictable inputs and outputs. AI agents are inherently unpredictable.
Disseqt’s premise is that we cannot trust the core models to police themselves. Even with RLHF and safety guardrails, a model can be manipulated via prompt injection or simply lose the plot during a complex chain of thought. If you give an agent the keys to your infrastructure, you are essentially hiring an intern who might suddenly decide to set the office on fire because they misinterpreted a memo.
The Insider Perspective
The fact that this team comes from OpenAI and Meta is significant. These are the people who saw how the sausage was made. They know that safety at the model level is a moving target and, frankly, an unsolved problem. By stepping outside the big lab environment to build a dedicated monitoring layer, they are signaling that safety should not be an internal feature of a model, but a separate piece of the stack.
From a founder’s perspective, this makes sense. If I am building an agentic platform, I don't want to rely solely on OpenAI's internal filters. I want an independent auditor—a watchdog that sits between the agent and the real world. Disseqt is trying to be that watchdog, monitoring the internal state of agents to catch "rogue" behavior before it executes.
The Problem with Rogue Agents
What does "rogue" actually mean in a business context? It is rarely a sci-fi scenario about sentient machines. It is usually much more mundane and damaging. It is an agent getting stuck in a recursive loop that racks up a $10,000 API bill in an hour. It is an agent accidentally leaking PII because it thought it was helping a customer. It is an agent making a commitment to a client that the company cannot legally fulfill.
Current solutions are mostly reactive. You look at the logs after the damage is done. The goal of new safety infrastructure is to be proactive—to analyze the intent and the trajectory of the agent's logic in real-time. This requires deep integration into the inference process, which is where Disseqt is focusing its efforts.
What This Means for Builders
If you are a developer building agentic workflows, you are probably feeling the tension between speed and safety. You want to ship the autonomous feature, but your CTO or legal team is terrified of the liability. The emergence of companies like Disseqt suggests a few things for your roadmap:
- Separation of Concerns: Don't bake your safety logic directly into your system prompts. It’s too fragile. Look toward middle-layer security that can handle validation independently.
- Audit Trails are Non-Negotiable: If you aren't logging every step of an agent's reasoning, you aren't building a product; you're building a liability.
- The "Human in the Loop" is Changing: We are moving from humans checking every output to humans managing high-level policy. We need tools that translate human intent into machine-enforceable constraints.
The Skeptic's View
While the pedigree of the Disseqt team is impressive, we have to ask if another layer of software can truly solve the unpredictability of a black-box model. If the safety layer itself is powered by AI, we are just adding more layers to the onion. There is a risk that safety startups become a bottleneck for performance or a new vector for failure themselves.
Furthermore, as the big labs like OpenAI and Anthropic release more robust versions of their own safety tools, specialized startups will have to prove they offer something significantly deeper than a wrapper around existing moderation APIs. The value will likely be in the "white-box" approach—actually understanding the internal activations of the model rather than just looking at the text it spits out.
Final Thoughts for Founders
The "Gold Rush" phase of AI is transitioning into the "Settlement" phase. In the settlement phase, we care about fences, laws, and insurance. Disseqt represents the first wave of AI insurance via infrastructure. They are betting that as agents become more powerful, our fear of them will grow faster than our desire for their utility.
The real challenge for builders isn't making an agent that can do the work; it's making an agent that knows when to stop.
We are heading toward a future where every company will have an "AI Policy Engine" that sits alongside their firewall. Whether Disseqt becomes the industry standard or just one of many attempts, the direction is clear. If you are building for the enterprise, start thinking about how you will police your agents today, before they do something you can't undo.
Read the original at Sifted →