Toronto Researchers Built an AI Worm That Hacked Two-Thirds of a Network Alone
Researchers at U of Toronto built an AI worm that exploited 73.8% of a test network autonomously, using stolen GPU compute and open-weight LLMs — no cloud APIs required.
Takeaways
- Researchers built an AI worm that autonomously exploited 73.8% of a 33-host test network in 7 days using only an open-weight LLM on a single GPU.
- The worm steals GPU compute from compromised machines to run its own LLM inference — no cloud APIs, no vendor controls, zero marginal cost per infection.
- It exploited three vulnerabilities disclosed in 2026 (after its training cutoff) by ingesting public advisories at runtime, converting them to working exploits.
- The key innovation is the agentic harness — a reasoning graph, hierarchical memory, and skill system that compensates for small-model limitations without fine-tuning.
- Exploitation succeeded in 44% of attempts; self-replication succeeded in 88% of successfully exploited hosts, reaching up to 7 generations.
- The paper withholds implementation details; code is restricted to vetted researchers, but the test environment will be open-sourced for defensive research.
A team of researchers from the University of Toronto, the Vector Institute, and the University of Cambridge has built and tested a proof-of-concept AI-driven computer worm that does something no malware has done before: it reasons about each machine it encounters, generates a custom attack strategy on the fly, and then replicates itself onto the compromised host to continue spreading. This is not a theoretical exercise. The worm ran for seven days, fully autonomously, and compromised nearly two-thirds of a simulated corporate network.
Why traditional worms are a solved problem , and why that's about to change
Unlike a traditional computer virus, which needs a human to be duped into opening an infected file, worms can infect devices entirely on their own by exploiting security vulnerabilities and replicating copies of themselves along the way. The classic defense is straightforward: identify the specific vulnerability the worm uses, patch it, and the spread stops. WannaCry and NotPetya both worked this way. They were devastating, but they were also finite , once the exploit was known, the fix was known.
An AI-driven worm breaks that model entirely. Researchers built and tested a proof-of-concept AI-driven worm that does not operate on a fixed list of exploits. Instead, it analyzes each target it encounters, reasons about how to attack it, and creates a strategy on the fly, all with the help of a small, free large language model running directly on machines it has already compromised. Patching one vulnerability does not stop it , the worm will simply try a different approach on the next target.
The parasitic compute trick that makes this economically terrifying
The most unsettling design choice is how the worm handles its own computational needs. Running an LLM requires serious hardware , typically a GPU with tens of gigabytes of memory. Rather than relying on a cloud API (which could be monitored, rate-limited, or shut down), the researchers built the worm around a single open-weight LLM that runs locally on a compromised GPU. Once the worm compromises a GPU-equipped server, it deploys a local copy of the model.
This creates a self-funding attack loop. Every new GPU-equipped machine the worm compromises becomes a new reasoning node, expanding the worm's capacity to attack more targets. The most significant aspect is not that the prototype exists, but how it works: it does not rely on cloud APIs, does not need a centralized infrastructure, and does not depend on OpenAI, Google, Anthropic, or any other vendor. The attacker's marginal cost per new infection is effectively zero , the victims are paying for their own compromise.