A 3AM ransomware affiliate breached a UK school estate in September 2026 in what Bitdefender researchers describe as the first hybrid human-plus-agentic operation they have documented in Europe. The attacker encrypted roughly 37 of nearly 50 machines reached from a single unmanaged host, and the intrusion shows a human operator setting goals while a model handled mechanical retries.
What Happened
The intrusion ran from 10 to 19 September 2026. Starting from an unmanaged host, the affiliate reached almost 50 machines across the school estate and successfully encrypted about 37 of them. Backup credentials were obtained on 10 September, and from 12-13 September the attacker moved laterally through WMI and PowerShell, ultimately deploying the 3AM encryptor. No data exfiltration or negotiation details have been confirmed publicly.
Three Kinds of AI Attack, and Why This One Is Hybrid
AI-involved attacks fall into three buckets. In an AI-generated attack, a model writes the malware itself, often in bulk (Bitdefender's "vibeware" research, March 2026). In an AI-assisted attack, human engineers build the tools and use AI to speed up the work, as assessed in the SilkParasite campaign (August 2026). In an AI-operated attack, AI works the intrusion itself: it chooses and writes commands, reads results and adjusts. This case does not fit a single bucket. Some parts show a human at the keyboard; others behave like an AI agent working throughout the intrusion.
Living Off the Land
With the exception of the encryptor itself, the whole operation used legitimate, built-in Windows tools: PowerShell, WMI, reg, netsh, rundll32, plus domain and scheduling utilities. A handful of public tools were brought in (an Impacket-style remote execution tool, Mimikatz, a published Veeam script), but nothing novel. In Bitdefender's analysis of roughly 700,000 security incidents, 84% of high-severity attacks involved LOTL binaries, and four of the five most abused binaries in that study (netsh, PowerShell, reg, rundll32) appear in this intrusion. The lesson: an AI-enabled attack need not be novel or creative to work.
Tells of an Agentic Loop
Several behaviors point to an agentic loop feeding each result back to a model and running what comes back:
- A call to the Veeam backup software's web interface used a port and login method that do not match vendor documentation, a mistake typical of writing from memory rather than the manual.
- Confirmation files, self-labeled output, and numbered revisions of the same script.
- On 13 September the attacker re-ran a PowerShell command unchanged 41 seconds later, changed a single call 49 seconds after that, then 47 seconds later abandoned PowerShell for
wmic. The complete rewrite arrived as fast as the minor changes before it. - Across nine retries, the delay before the next attempt ran 40-101 seconds regardless of change size, while moving to a new task took minutes. Commands before the retry sequence were simpler and sat two to sixteen minutes apart, the pace of someone reading output and deciding what to do next.
Joining the Domains
The clearest example is how the attacker joined separate Active Directory domains. Building domain trusts is mechanical, order-sensitive work that must be configured on both sides. The attacker built the trusts, ran diagnostic probes, and restarted the service maintaining the secure channel between domains. Bitdefender's reading, offered as interpretation rather than finding, is that this suits a model: repetitive configuration where the fix comes from reading an error and trying again. It is also work a defender can see, since a new trust between domains is rare and cheap to alert on.
What We Can and Cannot Show
Reports of "AI ransomware" blur three claims, supported to very different degrees here:
- A language model generated the commands — strongly supported by command content.
- An agentic loop chose each next action — behaves like one; retries arrive at a steady pace regardless of change size. A person copying and pasting between a shell and a chat window cannot be excluded.
- No human was involved in hands-on — not supported.
Bitdefender could not access the attacker's working machine, so nothing is documented about which model was used, whether it ran locally or remotely, or whether a person was copying and pasting between the implant and the model. What can be documented is behavior, and the behavior is that of an agentic loop. The hybrid reading is a working model for defenders to plan around, not an established fact.
The Bigger Picture
Bitdefender stresses this is one case, an anecdote, not a trend. Vibeware accounts for only about 0.07% of malware samples collected monthly, and most intrusions still look as they did before language models. The attack was not fast (even quickest retries took 40+ seconds) and not innovative (built-in tools, a few public ones, a commodity encryptor). What it shows is that the hybrid approach works and makes economic sense: a person sets goals, a model handles mechanical work. Hybrid attacks like this are expected to become more common in 2027.
Defending Against AI Cybersecurity Threats
For defenders, the practical takeaways: alert on rare changes such as new domain trusts; treat LOTL tooling (PowerShell, WMI, reg, netsh, rundll32) as high-signal in school and estate environments; monitor backup infrastructure like Veeam for anomalous API calls on unexpected ports; and assume AI-assisted retries will arrive at a steady, mechanical cadence. As AI cybersecurity threats evolve toward hybrid operations, the indicators are behavioral rather than novel tooling.
Sources
- BleepingComputer, "UK school estate hit by 3AM ransomware in suspected AI-assisted attack," https://bleepingcomputer.com/news/security/uk-school-estate-hit-by-3am-ransomware-in-suspected-ai-assisted-attack
- Bitdefender Business Insights, "3AM Hits European Schools: A Hybrid AI Ransomware Attack, Part Human, Part Agentic Loop," Marius Baciu, November 3, 2026, https://www.bitdefender.com/en-us/blog/businessinsights/3am-ransomware-hybrid-ai-attack
- Infosecurity Magazine, "3AM Ransomware Hybrid AI UK School," https://www.infosecurity-magazine.com/news/3am-ransomware-hybrid-ai-uk-school/
happened