Skip to content
BITBRIEF

Institutional research · AI · Cybersecurity · Digital assets

Vol. 01 · No. 22

What AI actually changes for attackers

Not new vulnerabilities. Who can attack, how many operations one crew can run, and what a defender finds in the payload.

In short

In every case we have reported, AI did not find a new kind of weakness. It lowered the expertise and time needed to exploit known ones — the NSA and FBI said exactly that of AI-assisted scripts against Siemens S7 controllers — which changes who can attack and how many attackers there are.

Scale is the concrete effect. South Korean police say one former penetration tester, using the ARTEX AI framework, breached more than 90 organisations and stole 8.2 million records, and call it the first Korean arrest for large-scale data theft using generative AI.

Much of the gain comes after the breach rather than before it. Talos found a group using agentic tools for exploit refinement, validation and documenting its own intrusions — the parts of an operation that take skilled time and do not scale.

AI-made code can give itself away. Unit 42 found conversational annotations and developer-reasoning comments inside payloads used against government Linux systems, from operators who also planted fake red-team markers to look like an authorised test.

Does AI find new vulnerabilities?

Not in anything we have reported. An advisory from the NSA, the FBI and other US agencies described actors using AI-assisted development to build exploitation scripts for Siemens S7 programmable logic controllers in the energy, water, agriculture and defence sectors, dressed up as legitimate monitoring tools. The advisory's own claim was that this dramatically reduces the expertise and time needed to produce a working exploitation script for industrial control systems. That is a statement about the supply of attackers, not about the sophistication of any single attack: nothing suggested AI found a new weakness in the controllers. Writing an exploit for a known one simply no longer requires years spent learning industrial protocols.

What changes is scale

The South Korean case makes it concrete. Police identified seven suspects in an operation that breached more than 90 organisations — schools, hospitals, large corporations, golf courses — and stole an estimated 8.2 million records, using the ARTEX AI penetration-testing framework in most of the attacks. The leader, a self-taught 32-year-old who had worked at a cybersecurity company, used conventional tools at first and adopted AI utilities in late August 2026 to expand the attacks. In 38 cases he demanded 360 million won not to leak the data. The police's own framing was that individuals who are not professional hacker organisations can now hack on a large scale.

Cisco Talos documented the same economics from inside an operation. The group it tracks as UAT-10147 used agentic tools not to write lures before an intrusion but after access: exploit refinement, adaptive troubleshooting, post-exploitation automation with validation steps, and documentation of its own intrusions, against a target list of roughly 170,000 URLs. Its conventional tooling was unremarkable and its privilege-escalation exploits were years old. What the AI bought was throughput — documentation and validation are the parts of an intrusion that take a skilled operator's time and do not scale, and automating them lets one crew run more operations against the same old vulnerabilities.

What does AI-made malware look like?

Sometimes like a draft. Unit 42 reported a campaign against government entities and state-affiliated infrastructure, active from October 2025 to April 2026, in which multiple deployed artefacts still contained AI-made conversational annotations and developer-reasoning comments — text addressed to a reader rather than code. The researchers read that as a sign the payloads were generated by a language model or executed by an automated agent. The same operators attempted to disable endpoint protection, erased logs and planted deceptive red-team markers so that the intrusion would look like an authorised security assessment.

CaseWhere AI was usedWhat it changed
NSA and FBI advisory, Siemens S7Writing exploitation scripts, disguised as monitoring toolsLess expertise and time needed for a working exploit
UAT-10147, Cisco TalosAfter access: refinement, validation, documentationOne crew runs more operations at once
CL-UNK-1211, Unit 42Generating or executing payloadsReasoning comments left in deployed code
ARTEX AI case, South KoreaPenetration-testing framework for most attacksOne ex-pentester, 90+ organisations, 8.2 million records
Where AI sat in the attacks we have reported

What should defenders change?

Mostly the volume they plan for, not the controls they rely on. The recommendations in the industrial advisory were the ones made for a decade — take controllers off the internet, patch what can be patched, turn on monitoring — and they still apply because the weaknesses exploited are the known ones. What has changed is how many actors can reach them and how quickly a single crew can work through a list.

  • Assume known, unpatched vulnerabilities will be found and used faster and by more people than before.
  • Treat exposure on the internet as the first problem: reconnaissance precedes effect, and AI shortens the step after it.
  • Look for the tells of generated code — conversational comments and reasoning left in payloads are an indicator in their own right.
  • Do not read a red-team marker as proof of an authorised test; one campaign planted them as camouflage.
  • Expect post-exploitation to move faster: the documented gains sit after the breach, in validation and automation.

Questions

Is AI helping attackers find new vulnerabilities?
Not in the cases we have reported. The gain is in exploiting known weaknesses with less expertise and time — the NSA and FBI said so of AI-assisted scripts against Siemens S7 controllers — which increases how many people can attack rather than how advanced any one attack is.
How much can one attacker do with AI tools?
In the South Korean case, police say one former penetration tester using the ARTEX AI framework breached more than 90 organisations and stole an estimated 8.2 million records, after adopting AI utilities in late August 2026 to expand the attacks.
Where do attackers use AI during an intrusion?
Often after the breach. Cisco Talos found a group using agentic tools for exploit refinement, validation, post-exploitation automation and documenting its own intrusions, which lets one crew run more operations at once.
Can you tell that malware was written by AI?
Sometimes. Unit 42 found payloads that still contained conversational annotations and developer-reasoning comments — text addressed to a reader rather than code — which it treats as a sign of generation by a language model or execution by an automated agent.

All guides