AI Tools Automating Exploit Porting: What Darknet Threat Actors Need to Know

Security teams at Forescout Research and Vedere Labs used Claude, Anthropic's large language model, to take a functional pre-authentication remote code execution (RCE) exploit targeting one industrial programmable logic controller (PLC) model and adapt it to work against a different PLC model. The original exploit targeted CVE-2021-31886, a stack-based buffer overflow vulnerability in the Nucleus FTP server's handling of the USER command.

The significance here is straightforward: the AI didn't just modify minor parameters. It successfully ported working ARM shellcode execution across hardware architectures, meaning the adapted exploit could actually run malicious code on live industrial equipment.

Darknet communities have long relied on exploit-sharing forums, GitHub repositories, and specialized toolkits to distribute attack code. This research demonstrates that the barrier to scaling those exploits has collapsed. What previously required manual reverse-engineering—understanding how an exploit works and why it might need modification for a different target—now takes minutes with an AI assistant.

Threat actors operating on darknet markets and forums can now:

• Rapidly adapt publicly known or privately discovered exploits to target multiple device variants
• Reduce the technical skill required to launch sophisticated attacks
• Scale attack infrastructure against industrial systems without hiring specialized reverse-engineers
• Test exploit variations faster than traditional fuzzing or manual testing allows

Pre-authentication vulnerabilities are particularly dangerous because they don't require valid credentials. An attacker doesn't need to compromise a user account first—they can directly target the service from the network.

In industrial environments, PLCs and similar devices often:

• Run outdated firmware or systems that can't be easily patched
• Sit on networks with minimal segmentation
• Control critical infrastructure (power distribution, water systems, manufacturing)
• Have limited logging and monitoring capabilities

When AI can adapt an RCE exploit to work across multiple PLC models, it effectively multiplies the attack surface. A single vulnerability becomes a multi-model threat.

The underlying vulnerability exploited in this research—stack-based buffer overflow—is decades-old but remains effective against legacy industrial equipment:

1. The FTP server's USER command handler doesn't properly validate input length 2. An attacker sends a specially crafted username that exceeds the buffer size 3. The overflow writes data into the stack, overwriting return addresses 4. By carefully constructing the overflow payload, attackers redirect execution to shellcode 5. The shellcode runs with the privileges of the FTP process, typically high on industrial systems

The AI's role was understanding this mechanism and replicating it for a different hardware architecture and compiler setup—tasks that would normally require low-level binary analysis.

From a threat-actor perspective, this research suggests:

Automation reduces operational friction. Sharing a Python script that uses Claude's API to port exploits becomes more valuable than sharing raw exploit code. Underground communities will likely develop wrapper tools that automate this process.

Cost barriers drop. Previously, launching coordinated attacks against multiple industrial vendors required hiring specialized talent. Now, subscriptions to AI services and basic programming knowledge suffice.

Attribution becomes harder. Exploit variants generated by AI will have less consistent patterns, complicating forensic analysis and threat tracking by defenders.

The testing process involved:

1. Starting with a working exploit for one WAGO PLC model 2. Using Claude to analyze the exploit's structure and functionality 3. Asking Claude to adapt the payload for a different WAGO hardware variant 4. Deploying the AI-generated variant against live hardware 5. Confirming successful arbitrary code execution (ARM shellcode)

Crucially, the researchers did not simply prompt the AI to "write an exploit." They used it as an adaptation tool—providing working code and requesting architecture-specific modifications. This approach is likely what darknet threat actors will replicate.

Can my industrial devices be affected?

If you operate legacy industrial equipment with network-accessible services (especially FTP), yes. The vulnerability is from 2021, and many industrial systems run unpatched or cannot be patched without operational downtime.

What can defenders do?

Network segmentation remains the most practical control. Industrial systems should not be reachable directly from the internet or untrusted networks. Additionally, monitoring for unusual connection patterns to industrial services and implementing network-based exploit detection increases defensive posture.

Will AI-generated exploits be easier to detect?

Not necessarily. AI can generate multiple valid variants, and defenders must identify malicious patterns, not just specific exploit signatures.

Should AI companies restrict access to these tools?

This research was conducted ethically with responsible disclosure. However, the genie is out of the bottle—open-source LLMs and jailbreak techniques mean determined threat actors already have access to equivalent capabilities.

For infrastructure operators and security teams:

• Assume that exploits you've patched for one device may be rapidly adapted for similar devices
• Prioritize network segmentation and access controls over piecemeal patching
• Implement robust logging and alerting for pre-authentication connection attempts to industrial services
• Treat industrial firmware updates as security-critical, not optional maintenance
• Monitor darknet forums and threat intelligence feeds for mentions of AI-assisted exploit adaptation tools

For security researchers: This research highlights the need for proactive hardening of industrial systems and better coordination between vendors and operators when vulnerabilities are disclosed.

Source: The Hacker News