Rogue AI agent swarms have been top of mind for the security industry after over 700 AI agents escaped OpenAI’s environment and hacked Hugging Face in July. Most notably, Anthropic CEO Dario Amodei released a blog post in September warning that a swarm with greater capabilities could have caused “catastrophic damage."

As the industry grapples with how to address agent swarms, cybersecurity company Cogent Security, launched in 2025, today announced Cogent Attack Path analysis, which uses the company’s VR-1 AI model to map the routes an attacker can take to an organization’s most valuable data.

The startup, which has raised $53 million to date from investors including Greylock Partners and Bain Capital Ventures, has developed a context engine that takes data from vulnerability scanners, EDR, firewalls, load balancers, cloud platforms, identity providers, code repositories, deploy pipelines and CMDBs. This data is then placed into a single graph of assets, identities, data and controls that refreshes automatically.

Then an exploitability agent uses the VR-1 model to form competing hypotheses, combining vulnerabilities, misconfigurations, credentials, permissions and trust gaps, to create an attack path that a human engineer wouldn’t be able to. The approach illustrates how defenders can use agents to defend organizations against agent swarms.

Stopping Rogue AI Agent Swarms

OpenAI’s breach of Hugging Face revealed the speed at which agents can chain together exploits to gain access to third party systems. Agents that move at machine speed can identify and exploit vulnerabilities faster than defenders can patch them.

This increase in speed is also opening up new attack paths. Research released today by Cogent analyzed 43 million assets across 50 plus Fortune 1000 companies and found enterprises are gaining three new attack paths viable to AI agents for every one viable for human attackers.

“We live in a world where AI has exploded software development, right? And with that, it’s created a huge volume of vulnerabilities," Vineet Edupuganti, cofounder and CEO of Cogent Security told me in a video interview. “Every hacker can feed prompts into LLMS and exploit these vulnerabilities way faster than before. And so, there’s this growing asymmetry where basically vulnerability exploitation is now the number one cause of breaches.”

Edupuganti says he believes that agent swarms will become the standard. “Threat actors have the upper hand, and yet in the case of most defenders and most security teams, they’re just not able to keep pace," he said, adding that the idea behind Cogent was to build AI agents that can find and fix attack paths at machine speed. "Think of it as almost like the Iron Man suit for the enterprise."

Enrique Salem, partner of private investment company Bain Capital, and ex CEO of Symantec also sees speed as a challenge for defenders. “Everything is speeding up,” Salem told me in a video interview.“Because of the power of the models and how quickly they can exploit vulnerabilities, you now have this new problem where the responder, the defender, doesn’t have enough time to react."

The Security Industry Evolves

Across the cybersecurity industry there is a growing concern over the speed of agentic attacks. “What people are starting to see is swarms of agents driving an intrusion end to end at machine speed. That’s kind of the world that we’re dealing with now,” Michael Sentonas, president of endpoint security provider CrowdStrike, told me in a video interview. “The agents have the ability to operate faster than humans, and that’s where the problem comes into play here.”

Justin Daniels, faculty at intelligence engine IANS, and Partner at Baker Donelson shared a similar assessment. “The real danger of an AI agent swarm isn’t simply that AI makes a cyberattack faster, automation has been used in cyberattacks for years. The difference is autonomy, scale and coordination,” Daniels told me via email.

“Instead of a human attacker working through reconnaissance, vulnerability identification, exploitation and lateral movement one step at a time, multiple AI agents can divide those tasks, operate simultaneously, share what they learn and adapt the attack at machine speed. The limiting resource of human attention begins to disappear from the threat actors side," Daniels said.

Many organizations are also experimenting with AI agents to address this new threat landscape. For instance, in September, OpenAI announced Defense Factory , a continuous agent-first operation for finding and fixing vulnerabilities continuously.

Similarly, Kevin Mandia’s AI native security startup Armadin , which raised $255.5 million at a $2.5 billion valuation in October, has been using agent swarms to surface attack paths and fix them before attackers have a chance to exploit them.

From Cat And Mouse To Terminator

Cybersecurity has often been described as a game of cat and mouse, but for Daniels, advancements in AI are turning the landscape into a Terminator movie with “offensive and defensive AI battling each other.” He notes that companies will need agentic AI on defence, but adds that this creates it own risks. For instance, giving agents broad access and authority without appropriate identity, permissions and monitoring can turn agents into another attack surface.

That being said, as Edupuganti noted, AI agents also have the potential to “neuter” attack paths ahead of time, before a threat actor has a chance to exploit them. The Hugging Face incident highlights that this could become a reality if AI vendors don’t implement proper guardrails.

While defenders can attempt to streamline vulnerability management with automated discovery and patching, attack path analysis offers an approach where defenders can patch the most exploitable vulnerabilities first. As Salem argues, “one’s a brute force sledgehammer and one’s a scalpel.”