Skip to content

53,600 AI-swarm attacks. Every one stopped in the kernel. 0 bytes leaked.See the evidence

Solution · AI-Swarm Attack Prevention

Stop the swarm before it spreads.

Attackers now send swarms of AI agents that try hundreds of variations in parallel. Solomon stops them where they act, inside the Linux kernel, in microseconds. Then agents investigate, and a person signs off on anything critical.

Recorded · Oct 2026NVIDIA DGX workstation and a live 2-node AWS EKS cluster (Intel + Graviton). Same swarm, same host: without Solomon every secret is gone in 17 ms; with Solomon, 0 bytes leave, whether 10, 50 or 100 agents attack.

Measured · same swarm, with and without

Same swarm. Nothing leaves.

Without Solomon the swarm emptied the host of its secrets in 17 ms. With Solomon, across 53,600 attempts in four environments, not one byte got out.1

1 Emulated AI-agent swarm evidence run, 1 Oct 2026, and the recorded swarm demo on the same host, Oct 2026. Each bar is the share of a batch of attempts that got secrets out: 20 batches of 100 without, 20 batches of 2,680 with. Canary data only, no real credentials. Method and every number.

With Solomon0 bytesleaked across 53,600 attempts. Every one denied or killed inside the kernel.

An emerging category

What an AI swarm is.

A swarm is many AI agents attacking together, rewriting their code as they go. It is not one smart intruder; it is a crowd that never gets tired.

  • Parallel agents

    • Many AI agents work the same target at once, each on a different technique
    • The whole attempt fits inside milliseconds
    • No human operator to slow down, tire or make noise
  • Self-rewriting code

    • The AI writes new code every round
    • Hashes, signatures and learned behaviours go stale before they ship
    • Every run looks like a first-seen sample
  • Relentless retries

    • A blocked agent tries a variation, then another
    • Failures cost the attacker nothing
    • Defenses that alert but do not block only add to the queue
  • Process spreading

    • Agents spawn helpers, shells and workers across the system
    • Killing one process leaves the rest running
    • Containment has to take the whole process tree at once

Why existing tools miss it

Signatures and cloud SIEMs arrive after it is over.

Signatures need a sample seen before. A swarm writes fresh code each round, so there is nothing to match.

Cloud SIEMs and alert-only agents do see the attack, but only after the events are shipped, parsed and read. By then the agents have already taken the keys, spawned helpers and moved on.

The only place fast enough is the system call itself.

Where each defense actsvs an AI swarm
  • Signature antivirusblind
    matches known samplesSwarm code is new every round.
  • Cloud SIEMtoo late
    queries shipped logsThe logs arrive after the swarm has finished.
  • Alert-only EDRtoo late
    detects and reportsA person reads the alert after the secrets are gone.
  • Solomon eBPF Clientin time
    decides on the system callThe call is denied or the tree killed in < 300 µs.

Layer 1 · the kernel acts

Every technique stopped in microseconds.

The Solomon eBPF Client runs policy as eBPF programs in the kernel. Credential and secret theft is denied inline, cloud-role theft through the metadata service is refused, and the attacker’s whole process tree is killed at once, so retries and helpers have nothing to run on.

It does not need a signature. It blocks the behaviour, so code that rewrites itself is stopped the same way the first time and the thousandth.

Emulated AI-agent swarm: 50–100 concurrent workers issuing real attacker system calls against canary data (no exploits, no real credentials) on Amazon Linux 2023, RHEL 9 and Ubuntu 20.04/22.04. Lab results from disposable AWS test hosts, a 2-node EKS cluster and an in-house production server. 6 of 8 kill-chain steps are blocked; shell spawn and covert tunnels are monitored.

Technique · MITRE ATT&CKLog scale · 1 µs to 100 ms
  • Steal cloud credentials fileT1552.0016–34 µs
  • Dump process secretsT155215–31 µs
  • Kill the attacker's process treecontainment31–130 µs
  • Steal cloud role credentials (IMDS)T1552.00555–200 µs
  • AI- or analyst-ordered kill, platform → hostend to end4–18 ms
  • Container escape killedT1611 · incl. process start13–56 ms

Medians. Source: Emulated AI-agent swarm evidence run, 1 Oct 2026. The end-to-end kill is an AI- or analyst-ordered kill travelling platform → host; container-escape timing includes process start.

Emulated AI-agent swarm evidence run, 1 Oct 2026
swarm attacks, every one stopped1
of secrets leaked2
concurrent attackers stopped; denial still ~0.2 ms3
Containment4
  1. 1Emulated AI-agent swarm evidence run, 1 Oct 2026: 53,600 attempts across four environments, 100% denied or killed.
  2. 2Emulated AI-agent swarm evidence run, 1 Oct 2026.
  3. 3Emulated AI-agent swarm evidence run, 1 Oct 2026: live production server, waves of 10, 50 and 100 attacking agents.
  4. 4Emulated AI-agent swarm evidence run, 1 Oct 2026: credential-file theft denied in 6–34 µs, process-secret dumps in 15–31 µs, attacker process tree killed in 31–130 µs, IMDS theft in 55–200 µs (medians).

Let’s stop the swarm. Together.

Run Solomon against your own workloads in a guided pilot. We’ll show you the kernel log, the investigation and the approval, on your infrastructure.