AI in offensive cybersecurity: from payloads to APT operations

AI as an accelerator, not an autonomous weapon

In recent months, the debate on artificial intelligence applied to cyberattacks has become more concrete. The most balanced reading, shared by much of the OT security research community, is that today AI acts mainly as an efficiency accelerator: it does not introduce radically new offensive capabilities, but it shortens development time, multiplies variants, and automates decisions that previously required manual work.

In other words, defenders are not yet facing fully autonomous threats. They are, however, already facing attacks that are faster, noisier, and more diverse from one another.

Where attackers are using AI

The integration of AI into malicious campaigns follows three main directions:

• Development and red-teaming assistance: models help write, refine, and port code across different languages.
• Direct payload generation: scripts and backdoors produced semi-automatically, often recognizable by unusually verbose comments.
• Runtime integration: malware that queries a language model during execution to make dynamic decisions.

A few documented examples make the picture tangible:

Threat	Use of AI
Slopoly	AI-generated PowerShell C2 backdoor for fast post-compromise access
Konni APT	PowerShell backdoor with extensive comments, a hint of automated generation
APT36	Mass production of variants by porting logic across Nim, Zig, Go, and .NET
PromptLock	Ransomware that queries an LLM at execution time for dynamic choices
PROMPTFLUX	VBScript dropper that regenerates obfuscated code in real time

Why signature-based detection is faltering

The critical point for defenders is this: AI-assisted development makes it possible to create many functionally identical but structurally different samples. By changing language, structure, and obfuscation at every iteration, attackers erode the effectiveness of detection based on static signatures.

AI does not generate new attack techniques, but it makes it cheap to produce infinite variants of the same one.

It must be said that many of these implementations remain experimental and still depend on traditional delivery and human intervention. The trajectory, however, is clear.

The MON5 angle: visibility and anomalies in OT environments

For those managing industrial and OT infrastructure, the practical consequence is that you cannot rely on static indicators alone. When the payload mutates at every execution, what stays stable is behavior: the sequence of actions, the execution context, the anomalous flows.

On this ground, OT defense has a few specific levers:

• Complete asset visibility: knowing which devices talk, to whom, and over which protocols is the basis for recognizing what is out of place. An unexpected API communication toward an LLM service from an unauthorized control system is already a strong signal.
• Correlating vulnerabilities with plant context: a CVE is not worth the same everywhere. Weighing it against exposure, process criticality, and network segmentation makes it possible to prioritize what really matters.
• Anomaly monitoring without stopping production: passive analysis of industrial traffic spots behavioral deviations without interfering with operational continuity, which in OT is non-negotiable.

In summary

Offensive AI is an operational reality, but its main effect is quantitative: more speed, more variants, more noise. The answer is not to chase every new signature, but to shift the center of gravity of defense toward behavior, context, and continuous visibility. For OT environments, where process integrity comes before everything, this is exactly the approach MON5 puts at the center.

Analysis and commentary by MON5 based on public-domain research and data from the OT/ICS sector.