Sweet Security Says It Can Block Rogue AI Agents Before They Act

Key points

• Sweet Security, founded in 2023, has raised $120 million and now claims its platform blocks misbehaving AI agents in real time.
• The new "Agentic AI Blocking" feature targets prompt injection attacks, where a hidden instruction hijacks an AI agent's behaviour, along with unauthorised data grabs.
• Sweet says its system analyses over one billion runtime events every day to learn what each agent is supposed to do, then enforces that limit.
• The company reports its platform already protects tens of thousands of cloud applications at Zoomd, a digital-advertising firm.
• Most competing security tools only alert a human after a problem occurs; Sweet's pitch is that the action gets stopped before it completes.

---

What is an AI agent, and why should you care?

An AI agent is a piece of software that acts on your behalf without needing you to click anything. Think of a digital assistant that can read your company emails, book meetings, query a database, and send replies, all on its own.

That autonomy is exactly the problem. An agent given broad access to a business's systems can touch sensitive information, customer records, financial data, internal passwords, without a human in the loop. If something goes wrong, it may already be too late by the time anyone notices.

Prompt injection, where an attacker hides a rogue instruction inside text the agent reads (a crafted email or document, for example), can silently redirect the agent to do something its owner never intended. It is, at its heart, the same trick as a classic SQL injection attack, where malicious instructions are smuggled inside a data field to manipulate a database. Same old idea; new target.

What does Sweet Security actually do?

It stops the bad action live, rather than writing it up in a report afterwards.

Sweet's platform sits at runtime, meaning it watches software as it runs, not just before or after. When an agent tries to call a tool it is not authorised to use, or attempts to send sensitive personal data outside the system, Sweet says it terminates that action immediately. The company calls this "Agentic AI Blocking."

The system builds a baseline of normal behaviour by watching over a billion events a day across its customers' environments. Once it knows what an agent is supposed to do, anything outside that boundary gets cut off.

Dror Kashti, Sweet's CEO, put it plainly: "When an agent reaches for data it shouldn't touch, Sweet stops the action red-handed and provides the team one line: here is what we found, and here is how we stopped it."

Capability	What it prevents
Unauthorised tool-call blocking	Agent using a system it was not given permission to access
Data-exfiltration blocking	Secrets, passwords, or personal data leaving through an agent
Prompt injection blocking	Hidden instructions steering the agent off its intended task
Session termination	Ending a rogue agent session before further harm occurs

Should businesses running AI tools pay attention?

Yes, particularly any organisation that has deployed AI assistants with access to customer data or internal records.

Sweet's pitch is timely. Security analysts have begun describing real-time enforcement as a baseline requirement for AI deployments, not a nice-to-have. The gap Sweet is targeting is real: most security platforms flag suspicious behaviour after the fact, and agents can move faster than a human can read an alert.

For ordinary employees, the practical takeaway is straightforward. If your company uses AI tools that can access customer records or company files, ask whoever manages your IT whether any controls exist on what those tools are allowed to do. Monitoring is not the same as blocking.

Sweet plans to demonstrate the platform live at Black Hat USA 2026, a major security-industry conference, in July.

Common questions

Does this affect me if my company doesn't use AI agents yet?

Not directly today, but AI agents are being added to common business software quickly. Keeping an eye on what your workplace tools are authorised to access is a sensible habit now.

How is this different from ordinary antivirus or firewall software?

Antivirus looks for known malicious files; a firewall controls which computers can talk to each other. Sweet's approach focuses specifically on the actions an AI agent takes at the moment it takes them, a layer those older tools were never designed to watch.