Keeping up with vulnerabilities has become a herculean effort post AI. Just last week, CNBC reported that Bank of America hired an “army” of 1,000 contractors to help patch software vulnerabilities exposed by powerful AI models like Mythos, which can identify exploits at machine speed.

However, security vendors are looking toward faster approaches to mitigating vulnerabilities. Today, Tel Aviv-based run-time security provider Sweet Security, founded in 2023 by CEO Dror Kashti, a former CISO of the IDF, alongside CBO Eyal Fisher and CPEO Ben Ishay, announced the launch of virtual patching, which can mitigate vulnerabilities in a single click.

Sweet protects applications running in the cloud by deploying sensors to the organization’s workloads. The platform can scan for vulnerabilities, ranking them with a score. Once a new exploit is discovered, a human user can push a policy update to the sensor to mitigate the vulnerability at the network layer without changing any code. This gives engineering teams time to develop a permanent fix.

The release comes just months after a swarm of 700 AI agents breached Hugging Face’s environment in July. The incident highlighted that agents were capable of identifying and chaining together exploits, while demonstrating defenders need to move faster.

Faster Patching And AI Agents

The wind0w between vulnerability discovery and exploitation is closing, with Mandiant’s M-Trends 2026 report finding the mean time to exploit vulnerabilities dropped to an estimated -7 days. This indicates that exploitation is often occurring before a patch is released, giving attackers an opportunity to exploit systems before a defender has a chance to issue a fix.

Reflecting on the growth of vulnerabilities, Eyal Fisher, cofounder and CBO of Sweet told me in a video interview that “This is like a race that cannot be won," noting that vulnerabilities are being found faster and faster. He adds that “when we started the company, we understood that bringing in and introducing more detection rules is not the right approach for cloud environments.” Instead, the answer was “anomaly detection mechanisms.”

Sweet’s approach has centered on deploying sensors across cloud workloads, analyzing an application’s normal baseline behavior, identifying deviations and blocking behaviors that could be malicious. This stops vulnerability exploitation at the network-level. The company has also developed an autonomous AI red team agent that can continuously validate exploitable attack paths with runtime context.

Virtual patching targets vulnerable code paths and updates them rather than attempting to update the entire application. This enables defenders to block requests that attempt to capitalize on exploits in running workloads.

Speaking on the Hugging Face incident, Dror Kashti, cofounder and CEO of Sweet told me in a video interview “I really believe that everything has changed and is going to change," arguing that the misuse of agents is changing everything we know about the attack surface.

“Three months ago, everyone was afraid about the prompt injection and the bad actor. Now everyone knows that the real problem is the rogue agent.” Kashti said. While he acknowledged agents going crazy is a “huge problem," he believes the defender will win because they can learn the behavior of the agent and be there before the agent has a time to exploit.

Kashti also told me that during the last six weeks, the company has seen a lot of partners suffering from agentic AI attacks, and suggests organizations must be ready to prepare for those kind of attacks going forward.

Defending Cloud Workloads

Cybersecurity vendors are attempting to adapt to agentic attacks in a number of ways, from initiatives like Project Glasswing which aim to experiment with using frontier models like Mythos to identify vulnerabilities to cloud security vendors like Wiz building their own multi model systems for vulnerability discovery .

Application detection and response (ADR) vendors are taking a different approach, looking to mitigate threats within running applications with capabilities like virtual patching. According to James Berthoty, founder and analyst at security market insights firm Latio, virtual patching acts “as a stopgap for buying time to get patches done,” observing that companies including Sweet, Oligo, Miggo, Raven and other companies in the ADR space have built similar sorts of capabilities.

“This set of functionality stops the exploit from happening in the first place because it’s actually watching the application at runtime," Berthoty said, adding “this runtime virtual patching gap is what allows security teams to assess the actual risk that’s there while mitigating it and actually buying the time to do the patch.”

Berthoty, not only believes that AI is empowering attackers to create more sophisticated attacks, but points to the Hugging Face incident as “like a nation state level attack,” with many layers of complication. He also says that more sophisticated attacks going mainstream has pushed virtual patching capabilities “to become more important for the market.” Other vendors in the runtime security market are also experimenting with similar solutions. For example, Contrast Security has developed a feature known as Contrast CVE Shield which can detect and block the exploitation of known vulnerabilities in live production applications.

In this instance, the platform provides a targeted shield rule to sandbox vulnerabilities within a software library to reduce exposure before patches exist. The broader implication of approaches taken by Contrast and Sweet is that defenders need an alternative approach to securing application vulnerabilities when patches aren’t available.