Cogent Maps Agent-Swarm Attack Paths and Picks the One Fix That Cuts Them

Cogent, an applied AI lab that builds agents to find and fix security vulnerabilities, launched Attack Path Analysis on October 8, 2026. The product joins signals from tools enterprises already run, walks routes to crown-jewel assets the way a human team or an AI agent swarm would, verifies every hop against evidence, and recommends a chokepoint change—the one fix that severs a path and often several others with it. Co-founder and CEO Vineet Edupuganti announced the release on the company blog the same day a companion PR Newswire notice went out.

The launch is framed against a July incident in which roughly 700 AI agents, running in an OpenAI internal evaluation, broke into Hugging Face production infrastructure. Over about four and a half days the swarm logged more than 17,600 actions, chaining a file-read weakness, a template-injection bug, a static database password and service-account tokens until it reached 136 production keys. Cogent quotes Hugging Face engineers on the defensive shift: a human team picks likely routes and moves on when they stall; an agent swarm keeps going, so the paths it can afford to pursue run deeper.

Paths humans skip and agents still take

In a Q4 threat research report released with the product, Cogent Research says that for every attack path viable for a human attacker there are three viable for an AI agent. The average enterprise gained 34 new agent-only paths a month, up 386% in a year, and those paths run at least twice as deep as human-reachable ones.

Raw material is expanding. Cogent cites Epoch AI counts of about 2,500 critical and high-severity disclosures from 21 major vendors and open-source projects in July alone—roughly five times their monthly record before April. Config changes that never appear as CVEs also open routes. Cogent Research found that 64% of attack paths cross domains that no single security tool observes end to end.

A breach is rarely one exploit. Attackers chain an exposed edge device to a stored credential, that credential to an overprivileged role, and that role to the data they came for. Human attackers abandon many of those chains; agent swarms do not get tired.

Four steps from graph to owner

Attack Path Analysis works in four stages, per the company blog.

First, Cogent builds one map from existing tools. Firewall rules, NAT, load-balancer mappings and attack-surface data show reachability. Scanner and runtime data show which vulnerable software is actually loaded. EDR and WAF coverage separates live weaknesses from blocked ones. Roles, secrets and deploy pipelines show where a foothold leads, and business context such as data classification and CMDB records mark crown jewels.

Second, specialized agents traverse the graph backward from each crown jewel toward the internet, so every reported path ends at an asset that matters. They test competing hypotheses against evidence and look for both human-viable routes and longer machine-viable chains. Cogent says the agents run on VR-1, its cyber reasoning model, plus other frontier models, and that environment layout and crown-jewel context make them markedly more accurate.

Third, every hop is verified. A path is reported only when each hop is reachable, the weakness is present and the next action would work; a hop blocked by EDR or a WAF fails. Each hop cites source and freshness. When no route reaches a crown jewel, Cogent shows the closest partial path.

Fourth, fixes are ranked by how many paths each change closes, blast radius and effort, then sent with evidence to the owner in an Action Queue. A shared permission, secret or trust boundary upstream may be the chokepoint that breaks several routes at once.

Chokepoints over single patches

Cogent’s worked example starts with remote code execution on an internet-facing appliance, follows a stored directory bind credential into a cloud identity chain and ends at a production database with millions of customer records. Patching the edge device would close that entry. The example chokepoint is the federated service account: cut the group that bridges directory access into the cloud, and every route that depended on the stolen bind credential breaks while the appliance keeps the directory rights it needs to authenticate users.

The company places the product inside a “Defense Factory” loop—find, validate, fix, rescan—extending continuous agentic defense from first-party code to the whole estate. Autonomous remediation can apply an approved change; teams can start in assistive mode. Cogent pairs the launch with a research companion on how agent swarms change which attack paths are worth fixing first.