U.S. cybersecurity company Cogent introduced Attack Path Analysis on Thursday, October 8, 2026, a product designed for a kind of threat that until recently seemed theoretical: swarms of artificial intelligence agents able to attack systems continuously and in coordination. The tool in turn uses AI agents to find the routes that would lead an attacker to an organization’s most valuable assets and to identify the change that severs them.
The context the company lays out is unsettling. Its CEO and co-founder, Vineet Edupuganti, recalled in the announcement that in July a swarm of roughly 700 AI agents broke into the production infrastructure of Hugging Face, the leading platform for sharing artificial intelligence models. According to Cogent’s account, those agents took some 17,600 actions, harvested credentials and moved laterally across its systems, at a scale no human hacking team could match.
Edupuganti warns that models with those capabilities are now available to anyone. He cites as an example GLM-5.3, the open-weight model Z.ai released in late August, and argues, citing an Anthropic analysis, that it builds working end-to-end exploits nearly as often as Claude Mythos Preview. According to Cogent, that model’s safeguards can be removed for a few thousand dollars of compute.
Alongside the product, the company published its quarterly threat report, titled “Beyond the Human Horizon.” Cogent analyzed 43 million assets and 1.2 billion vulnerability, misconfiguration and identity findings across more than 50 Fortune 1000 organizations. The main conclusion is that for every attack path viable for a human attacker, there are three viable for an AI agent.
The report’s figures show how fast things are changing. In August 2026, the average enterprise gained 45 new attack paths to its critical assets: 11 viable for human attackers and 34 viable only for AI agents. Human-viable paths grew 38% in a year, from 8 to 11 a month; machine-only paths surged 386%, from 7 to 34, as the offensive capability of models advanced.
The logic is simple. A breach is rarely a single flaw: attackers chain an exposed device to a stored credential, that credential to an overprivileged role and that role to the data they want. Humans abandon most of those chains because they are too long or uncertain. AI agents do not get tired, work through every route methodically and keep going where a person would give up. That is why Cogent distinguishes between human-viable and machine-viable paths.
The company also notes that 64% of attack paths cross domains that no single security tool fully observes. The vulnerability scanner sees the vulnerable device, identity tools see the overprivileged account and the cloud security platform sees the exposed database, but none of them sees the route that runs through all three. Cogent adds that fewer than 1% of isolated findings are exploitable where they sit.
Attack Path Analysis works in four steps. First, it joins the signals from the tools a company already uses into a single map, from firewall rules to roles, secrets and data classifications. Next, agents specialized in exploitability traverse that map from each critical asset back toward the internet, forming hypotheses about how an attacker could get in. The agents run on VR-1, Cogent’s cybersecurity reasoning model, and on other frontier models.
The third step is verification: at each hop, the system checks that the target is reachable, that the weakness is present and that the attacker’s next action would work, and it discards hops blocked by other defenses. Each hop cites its source and how fresh it is. Finally, the tool looks for the “chokepoint,” the single change that severs a path and often others at the same time, and sends it to the relevant owner as a concrete work item.
Cogent illustrates the method with a seven-hop example. A remote code execution flaw on an internet-facing F5 BIG-IP device allows an attacker to steal an Active Directory credential; that account is federated into AWS, which allows assuming roles, deploying code to a payments service and reading a database holding nine million customer records. Patching the F5 device would close only that entry point; according to Cogent, the most effective change is removing the service account from the group that connects it to the cloud.
The company frames its offering around the idea of a “Defense Factory,” a term OpenAI used in September to describe a continuous loop in which AI agents find, validate and fix vulnerabilities in its own code. Cogent says it brings that loop to the entire enterprise estate, with an autonomous remediation feature that applies changes once the customer authorizes them and only marks a path as closed after confirming the fix held.
The announcement reflects a growing fear in the industry: that AI agents will change the economics of cyberattacks. According to Epoch AI data cited by Cogent, about 2,500 critical and high-severity vulnerabilities were recorded in July across 21 major vendors and open-source projects, roughly five times the monthly record before April. If defenders do not automate their work, the company warns, they will be overwhelmed.
The figures should nonetheless be read with caution. Both the report and the product come from a company that sells solutions to that very problem, and the numbers on machine-viable paths depend on its own methodology. Even so, the trend they describe matches warnings from labs and security agencies. As Edupuganti puts it, someone will eventually get into a company’s network; the goal is that when they do, every route to its most valuable assets has already been severed.
