Once inside a network, attackers can spread to other systems in under 30 minutes, and sometimes in under 30 seconds. Most banks, utilities, and other organizations wouldn’t notice an intrusion in that time, let alone stop it. At INT21, we believe ultrafast inference and collaborative, self-improving agent swarms will be essential to defending that infrastructure.
Key takeaways #
AI is accelerating attacks: hackers now spread through a network in 29 minutes on average, with 27 seconds being the fastest time. #
Defense needs two things: collaborative agent swarms that chase every lead in parallel, and ultrafast inference so that each agent reasons more rapidly. #
INT21’s agent swarms have already built a runtime for a brand-new model in one hour, which ran up to 63% quicker than the officially supported option. INT21 can help cybersecurity teams build faster, more scalable defenses that outpace attackers.
Attackers now move at machine speed #
In July, researchers at cloud security firm Sysdig documented what they deemed the first agentic ransomware attack. An AI agent broke into a company’s systems, fired off more than 600 commands, and locked up more than 1,300 database records for ransom. When its attempt to give itself admin access failed, it diagnosed the error and found another way in within 31 seconds.
Such AI-enabled attacks are escalating rapidly. According to CrowdStrike, they increased by 89% last year, and while attacks spread within minutes, Mandiant found that the median time to detect an intrusion is 14 days. At INT21, we believe that spotting and stopping intrusions before they do damage requires ultrafast inference and collaborative agent swarms.
Why cyber defense needs speed and swarms #
Faster inference used to be hard to justify against cost, because a person reading an AI’s answer can only read so fast. AI agents remove the person from each step, and speed becomes the difference between catching an attack and missing it.
Ultrafast inference shortens reasoning cycles, helping individual agents investigate and respond sooner. #
Collaborative agent swarms expand coverage, allowing many agents to work in parallel and coordinate their response.
Every step an agent takes waits on inference, the time a model needs to produce its answer. Those steps include reading logs, forming a hypothesis, and checking another system. A single investigation can run dozens of steps in sequence, so even small delays compound. To put this into perspective, at 1,300 tokens per second, an agent could complete a 50-step investigation in about a minute, fast enough to run it more than 20 times before the average attacker reaches a second system.
An agent swarm is many specialized agents working toward one goal and sharing what they find. For example, one group of agents could trace a stolen password, another could inspect every system it touched, and a third could look for the same pattern elsewhere. Ultrafast inference and agent swarms depend on each other: a single fast agent can only follow one trail at a time, and a large swarm running on slow inference reaches every trail after the attack has already moved on. The combination of ultrafast inference and agent swarms gives defenders the best chance to contain a breach before it gains ground.
What we’ve built #
INT21’s Inference Engine Factory uses agent swarms to generate optimized engines across models. On standard 8×B200 systems, our engines deliver:
DeepSeek v4.1 Flash: more than 1,700 tokens per second #
GLM 5.3: more than 690 tokens per second
More examples can be found in our Inference Engine Factory. Behind this factory are swarms of thousands of agents, now in their second generation. The swarms build inference engines, and the underlying technology is general purpose. Because the swarms are self-improving, each engine they build makes the next one faster. Our Drama 3 inference runtime, built in one hour, is one example of what they can produce.
Giving defenders the speed advantage #
When we enabled self-replication in INT21’s agent swarms in a controlled environment, we witnessed its resilience. That experience highlighted a troubling possibility: an attacker could build a swarm that moves across compromised machines, using each new foothold to expand its reach.
Defenders need the same adaptability within their own systems: agents that deploy and recover across the computing resources they are authorized to use, share what they find, and coordinate a response. Ultrafast inference can help them act within the shrinking window before an attack spreads.
The same agent swarms that build INT21’s ultrafast inference engines can help cybersecurity companies build faster, more scalable defenses. Our goal is to put these capabilities in the hands of the teams protecting the systems we all rely on.
Security teams interested in working with INT21 can reach us at info@int21.ai.