LAS VEGAS — July 29, 2026 — Sweet Security today announced Agentic AI Blocking, a proactive runtime enforcement capability that stops rogue AI agents in live production before they can execute unauthorized actions. As enterprises race to deploy autonomous agents that interact with sensitive data and systems, the security industry has largely relied on detection and alerts — a model that leaves organizations reacting after damage has already occurred. Sweet Security’s new capability flips that paradigm, blocking malicious or misaligned agent behavior in real time using a continuous learning loop that understands what each agent is intended to do. With eighty percent of the world’s businesses already operating as AI enterprises, the need for runtime enforcement rather than post-incident notification has become acute.
The New Frontier of AI Security: Runtime Blocking for Rogue Agents
AI agents have fundamentally changed the enterprise attack surface. These autonomous software entities take on identities, reach sensitive data, and act on their own. They operate at machine speed, making and executing decisions in milliseconds. Yet the security tools designed to protect them have remained stuck in a detection-and-alert paradigm: they observe suspicious behavior, generate an alert, and wait for a human team to respond. By the time the team reads the alert, the agent has already acted — potentially exfiltrating secrets, modifying critical data, or executing unauthorized tool calls.
Sweet Security’s Agentic AI Blocking addresses this gap head-on. Instead of detecting and alerting, it intercepts and terminates rogue behavior at runtime, before any damage occurs. The company’s CEO and co-founder, Dror Kashti, framed the challenge succinctly: “AI has collapsed the cost of attack, and it has put autonomous software inside the enterprise. Someone has to decide, in the moment, what an agent is allowed to do.”
How Agentic AI Blocking Works at Runtime
What is Agentic AI Blocking?
Agentic AI Blocking is a runtime enforcement technology that stops AI agents from performing unauthorized actions in real time. It terminates tool calls, sessions, and data flows that deviate from intended behavior, blocks the exfiltration of secrets or personally identifiable information, and prevents prompt injections from steering agents off course — all before any harm occurs. The system relies on continuous self-learning rather than static rules.
The core engine behind this capability is the Sweet Learning Loop, which continuously Attacks, Fixes, and Defends. Sweet’s runtime reasoning layer analyzes over one billion runtime events every day, learning what every application and agent is intended to do. This deep contextual understanding allows Sweet to set a simple but powerful bar: “do exactly what your creator intended and nothing more.” Because the system knows the expected behavior, it can enforce decisively without breaking live production or generating false positives that would erode trust.
This approach stands in sharp contrast to traditional security models that rely on predefined signatures or behavioral baselines that take weeks to tune. Sweet’s automated learning adapts as agents evolve, ensuring that protection remains current without manual intervention.
Key Capabilities: Stopping Unauthorized Actions, Data Leakage, and Prompt Injections
Agentic AI Blocking delivers three specific enforcement capabilities at runtime:
- Terminates unauthorized tool calls and sessions – When an agent attempts to invoke a tool, API, or system function that falls outside its designated scope, Sweet blocks the call instantly. The session is terminated without allowing any data to pass.
- Stops secrets, PII, and sensitive data from leaving through an agent – Agents often have access to credentials, customer information, or intellectual property. Sweet inspects outbound data flows and prevents exfiltration, even if the agent’s actions appear legitimate on the surface.
- Blocks prompt injections live – Prompt injection attacks attempt to trick an agent into ignoring its instructions and executing malicious commands. Sweet identifies these attacks as they happen and blocks them before the agent deviates from its intended behavior.
Each of these capabilities operates autonomously, requiring no human decision-making at the moment of enforcement. The security team receives a concise report: here is what we found, and here is how we stopped it. Nothing bad happened, and nothing is waiting in a queue.
The Sweet Learning Loop: Continuous Self-Improvement Through Runtime Reasoning
The intelligence behind Agentic AI Blocking comes from Sweet’s proprietary runtime reasoning layer and the Sweet Learning Loop. This loop operates in three phases: Attack, Fix, and Defend. In the Attack phase, the system continuously probes for vulnerabilities and deviations in agent behavior. The Fix phase automatically adjusts the enforcement policies based on what was learned. The Defend phase applies those policies in real time, blocking any actions that violate the intended behavior.
This self-improving cycle is critical because AI agents are not static. Their behavior changes as they interact with new data, tools, and environments. A rule-based system that blocks a specific API call today may miss a new attack vector tomorrow. Sweet’s runtime reasoning layer, which ingests over one billion runtime events daily, allows the system to detect subtle shifts in agent behavior that indicate a compromise or misuse.
The bar Sweet uses — “do exactly what your creator intended and nothing more” — is deliberately simple. It does not require complex policy documents or manual whitelisting. Instead, it relies on the system’s ability to learn the baseline from observing normal agent operations. This makes it scalable across thousands of agents and tens of thousands of applications, as evidenced by Sweet’s deployment at companies like Zoomd.
Market Convergence: Analysts and Enterprises Demand Proactive Enforcement
The market is converging on the same conclusion that Sweet Security has been advocating. Leading industry analysts now describe runtime inspection and enforcement as a mandatory capability for securing AI agents. They expect the winners in agent security to be products that automatically prevent risky agent behavior rather than surfacing more dashboards and alerts that security teams cannot keep up with.
Sweet goes further by enforcing across both cloud and AI in a single platform. This unified approach is significant because AI agents do not operate in isolation; they interact with cloud infrastructure, databases, and third-party services. A security tool that protects agents but ignores the underlying cloud environment leaves critical gaps. Sweet’s platform covers both, providing consistent enforcement from the runtime of the agent to the runtime of the cloud applications it accesses.
The company’s funding and backing — $120 million from Evolution Equity Partners, Munich Re Ventures, Glilot Capital Partners, and Key1 Capital — reflects investor confidence in this proactive approach. Founded in 2023 by veterans of the IDF’s most elite cyber units, Sweet Security is built on a thesis that the security industry’s reliance on detection is not just insufficient but dangerous in the age of autonomous AI.
Customer Proof Point: Zoomd Blocks Risk at Scale
Sweet already enforces protection at extreme scale, including at Zoomd, where the platform protects tens of thousands of cloud applications in an environment where disruption is measured in market impact. Niv Sharoni, CTO at Zoomd, described the difference Agentic AI Blocking makes: “With most tools, you find out about bad behavior in a report after the damage is done. With Sweet, it’s blocked in runtime, the moment it happens. That’s the difference between monitoring risk and actually removing it.”
Zoomd’s experience underscores a critical point: in high-velocity environments, traditional alerting creates a backlog that security teams cannot clear. Automated blocking removes the need for triage, freeing professionals to focus on strategic threats rather than reviewing logs of already-compromised systems.
Demonstration and Availability at Black Hat USA 2026
Sweet Security will demonstrate Agentic AI Blocking live at Black Hat USA 2026, Booth 5721. Attendees can see how the system stops rogue agents in real time, from unauthorized tool calls to prompt injection attacks. The demonstration will highlight the Sweet Learning Loop in action, showing how the system adapts to new agent behaviors without manual configuration. More information is available at hi.sweet.security/black-hat-2026.
About Sweet Security
Sweet Security delivers proactive runtime enforcement for cloud and AI. While the rest of the industry detects and alerts, Sweet blocks bad behavior before anything bad happens, where intent turns into action: in runtime. Through the Sweet Learning Loop, organizations continuously Attack, Fix, and Defend, driving every application and AI agent toward maximal immunity. Founded in 2023 by veterans of the IDF’s most elite cyber units and backed by $120 million from Evolution Equity Partners, Munich Re Ventures, Glilot Capital Partners, and Key1 Capital, Sweet protects the world’s most demanding enterprises. Learn more at www.sweet.security.
The arrival of Agentic AI Blocking marks a fundamental shift in how enterprises can approach AI security. Instead of hoping that detection tools will catch a rogue agent before it causes irreversible damage, organizations can now deploy autonomous protection that acts in the moment. As AI agents become more autonomous, more connected, and more capable, the ability to enforce intended behavior at runtime will separate companies that adopt AI with confidence from those that are forced to operate in constant fear of the next breach. Sweet Security’s technology puts that confidence within reach — not through more alerts, but through decisive, automated enforcement.
