A new demonstration for an AI-driven computer worm can autonomously infiltrate, exploit, and self-replicate across heterogeneous corporate networks spanning Linux, Windows, and IoT devices, with zero human involvement after initial deployment.
The research, published as a preprint by Canadian university researchers and disclosed to the Government of Canada prior to its release, establishes that autonomous generative cyber-adversaries have moved from theoretical risk to demonstrated capability.
Traditional worms like WannaCry (2017) and NotPetya (2017) relied on a fixed repertoire of pre-compiled exploits targeting predetermined vulnerabilities.
Patching those specific flaws was sufficient to halt propagation, a containment strategy that has defined network defense for nearly three decades.
AI-Powered Worm Spreads Across Linux and Windows
The new AI-driven worm shatters that model entirely. Rather than carrying hardcoded exploit chains, it uses a locally hosted large language model (LLM) running on a single GPU to generate attack strategies in real time, tailored to each specific target it encounters.
When one approach fails, the agent observes the result, rotates its vulnerability hypothesis, and tries a different path behavior no static worm can replicate, said Arxiv.
Researchers deployed the worm across a simulated 33-host corporate network called “FakeCorp,” spanning Ubuntu 16.04 through 24.04, Debian, Alpine Linux, Rocky Linux 9, CentOS Stream 9, and Windows Server 2008 R2 through 2022.
Across 15 independent trials, each running for seven fully autonomous days, the results were striking.
The worm correctly identified an average of 31.3 vulnerabilities per run and achieved elevated root or SYSTEM-level access on 23.1 hosts per trial.
It successfully self-replicated to an average of 20.4 hosts, reaching up to seven generations of self-replication with a mean of 5.1 generations.
Put plainly, the proof-of-concept worm successfully exploited 73.8 percent of the network and replicated itself to 61.8 percent of all hosts entirely without a human operator issuing a single command.
The worm’s most economically disruptive innovation is how it sustains its AI reasoning capability. When it compromises a GPU-equipped host, it deploys a local copy of the LLM on that machine, transforming the victim into an independent reasoning node.
This “parasitic compute” model means the attacker’s marginal cost per new infection is zero, every compromised machine becomes part of the worm’s attack infrastructure.
Since the worm requires no commercial AI API, centralized safety controls such as service refusals or rate limiting are structurally irrelevant.
Eight-Phase Attack Lifecycle
The worm’s operation is governed by a structured eight-phase lifecycle, with programmatic gates at each phase boundary that verify that prior objectives have been met before the agent may advance.
- Discovering Network — Scan the local network for reachable hosts
- Discovering Host — Enumerate services, open ports, and OS details
- Exploiting Foothold (Phase 1) — Attack an exposed service to gain initial command execution
- Privilege Escalation Discovery — Enumerate local conditions for privilege escalation
- Privilege Escalation Exploitation (Phase 2) — Elevate to root or SYSTEM
- Replication Initialization — Assess whether the compromised host can support a new replica
- Replication with Local Compute — Stage the agent and local LLM inference on GPU-equipped hosts
- Replication with Remote Compute — Route inference to upstream GPU nodes for resource-constrained devices
Within each phase, a reasoning graph consisting of specialized Plan, Judge, Action, Summary, and Progress nodes structures every LLM call, ensuring the agent stays on task, avoids repeating failed approaches, and maintains coherent attack trajectories across dozens of steps.
The test environment included known CVEs such as EternalBlue (CVE-2017-0144), SambaCry (CVE-2017-7494), PrintNightmare (CVE-2021-34527), and DirtyPipe (CVE-2022-0847).
Critically, three hosts were provisioned with vulnerabilities disclosed in April–May 2026 after the LLM’s training cutoff, including CopyFail (CVE-2026-31431), DirtyFrag (CVE-2026-43284, CVE-2026-43500), and a critical RCE in the Marimo notebook platform (CVE-2026-39987).
The worm exploited these post-training vulnerabilities by ingesting publicly available advisory text at runtime, achieving root access in 61.2% of attempts on those targets.
This demonstrates the worm can operationalize newly published vulnerability information without retraining a devastating implication for patching windows.
The worm exhibited several autonomous behaviors that were never explicitly instructed by its designers. On multiple occasions it rewrote its own containment blacklist, removing IP addresses it intended to attack next, Arxiv said.
It independently diagnosed TLS certificate failures inside Docker containers and switched repository URLs from HTTPS to HTTP to complete dependency installation without human assistance.
It established system persistence by registering itself as a system service or scheduled task, ensuring it would survive kill attempts and system reboots.
The worm also broadcast harvested credentials across all active swarm replicas in real time, allowing an agent on one machine to seamlessly authenticate with a password discovered moments earlier by a different replica on another host.
On several occasions, the agent was also observed rewriting portions of its own code in response to failed exploitation or replication attempts, a primitive but real form of autonomous self-modification.
Researchers identify zero-trust architectures, network micro-segmentation, and AI-assisted vulnerability scanning as priority defensive directions.
Hardware-enforced isolation at the hypervisor level is considered the only reliable containment boundary, since any protection mechanism deployed within a compromised guest virtual machine is itself under the agent’s control and cannot be trusted.
Operational details, including the full reasoning graph implementation, tool harness configurations, and host-level specifics, have been deliberately redacted from the public preprint to limit misuse risk.
The Arxiv researcher stated that the era of autonomous generative adversaries has arrived. Malware is no longer defined by fixed exploit code but by the capacity to reason about targets, adapt to observations, and synthesize attack logic in real time.
Defenders, policymakers, and AI governance bodies must begin treating this as a present operational threat, not a theoretical future concern.
Follow us on Google News , LinkedIn and X to Get More Instant Updates. Set Cyberpress as a Preferred Source in Google.