A PowerShell Script That Screamed ‘AI’
On June 3, a threat actor broke into a corporate network. They didn’t use a polished exploit kit or a dark-market trojan. They used a PowerShell script that an LLM had vibe-coded — built line by line through plain-English prompts, with errors pasted in until it finally ran.
Security firm Huntress recovered the script from the incident and published its analysis on July 8. The file’s name alone gave it away: “100% Working AD Information Gathering Script – FULLY FIXED”. That phrasing, Huntress noted, is a dead giveaway of a back-and-forth with a large language model — a human copying error messages back into the chat until the AI spat out something that compiled.
The attacker didn’t even bother to edit out the placeholder server name the LLM had invented as an example. It was still in the code, untouched.
What the Vibe-Coded Malware Actually Did
Once executed, the script set out to map the target’s Active Directory environment. It harvested user accounts, computer names, security groups, and trust relationships, dumping everything into neatly formatted spreadsheets. Then — and this is the part that caught researchers off guard — the LLM had apparently added its own flourish: a tidy HTML report summarizing the stolen data.
“The script was over-engineered in the way only an AI would produce it,” Huntress wrote. A human coder would pick one method to find the domain controller. This script had five separate fallback methods, each one more convoluted than the last. It also loved colorful console output — another AI hallmark.
The AI Fingerprints Were Everywhere
Huntress called the script a “case study in how criminals are weaponizing AI.” The telltale signs included:
- Left-behind comments from the LLM’s example code
- Unnecessary complexity in logic branches
- A default server name that was never replaced
- Excessive error-handling loops that no human would write
The Same Old Smash-and-Grab, Just Faster
Despite the novelty of the tool, Huntress was careful to say: AI isn’t changing the game fundamentally. The intrusion followed a classic playbook. The attacker logged in over RDP with stolen credentials, dropped the script into a common Windows folder, and ran it to scout the terrain.
For exfiltration, they used legitimate cloud tools — s5cmd and SharpShares — to siphon the data out. No custom backdoors. No sophisticated C2 infrastructure. Just a familiar smash-and-grab, accelerated by an AI-generated reconnaissance tool.
Why Signature Detection Fails Against Vibe-Coded Malware
Here’s the real headache for defenders: that script was one of a kind. It had never been seen before, and it will almost certainly never appear again in the exact same form. The file hashes and static signatures that antivirus products rely on were completely useless.
“Vibe coding lowers the barrier to entry for cybercrime, allowing unsophisticated actors to generate highly capable, evasive tooling on the fly,” Huntress warned. Even a mediocre attacker can now spin up bespoke, one-off tools with nothing more than a browser and a clear prompt.
What Defenders Should Do Instead
Huntress’s recommendation is blunt: abandon signature-based thinking. Behavioral analytics — watching what a process does rather than what file it is — catch the underlying actions that no LLM can hide. The script still had to enumerate Active Directory. It still had to reach out to the domain controller. It still wrote spreadsheets to disk. Those behaviors are hard to mask, regardless of how the code was generated.
The Takeaway for Security Teams
This isn’t a warning about some future threat. It’s a post-mortem of an attack that already happened. The era of AI-generated, vibe-coded malware is here, and it’s handing unsophisticated criminals the ability to produce custom tooling on demand.
The code may be messy, over-engineered, and littered with AI hallmarks like orphaned comments and placeholder server names. But the threat it poses is very real. Defenders need to shift their focus from “what file is this?” to “what is this process doing?” — because the AI can write the script, but it can’t hide the behavior.
