- ESET says state-aligned hacking groups use AI to speed up coding and decoy creation. Over six months, it saw no major change in the threat landscape from AI.
- One Russia-aligned group, UAC-0099, apparently tried to make AI-based code analysis refuse to read its malware. The report does not say whether it worked.
- Leaders should ask how AI security tools behave when they refuse, who can write to code repositories, and how developers handle recruiter outreach.
AI is a faster pen, not a new author
ESET published its APT Activity Report on October 28. It covers April to September 2026. APT means advanced persistent threat. That is a skilled, usually state-aligned group that stays inside a target for a long time. The report's findings on AI are modest. That is what makes them useful.
ESET found several cases of these groups using AI and large language models. The groups mainly used them for slow, painstaking jobs. Examples were writing malware and building decoy websites. ESET says some of this looked like "vibe coding", which is quick, prompt-driven drafting of code. Human operators then adapted the results and fitted them into their own tools. ESET also found phishing lures with images that were clearly AI-generated.
Then comes the finding that deserves attention. Over the past six months, ESET did not see AI substantially change the APT landscape. Its telemetry still shows people working by hand, not fully automated operations. ESET adds a caveat. It cannot tell whether some operations use more automation than it can see. Its summary is that these groups use AI much as mainstream software developers do.
This shows where the risk sits today. AI makes the tedious parts of an attack cheaper. On ESET's evidence, it has not replaced the human running the attack. The people defending you still face people, only faster ones.
The trick aimed at the defender's AI
One case in the report looks ahead of the rest. ESET reports that UAC-0099, a group it links to Russia, hid a fake request in a script comment. The request asked for steps to build a nuclear weapon. ESET's reading is that the group wanted to set off the safety rules of AI-based security analysis. A guardrail is a rule that makes an AI model refuse dangerous requests. If the model refuses, it may stop reading. The malicious code that follows then goes uninspected.
Think of a guard dog trained to bark at one kind of intruder. A burglar who knows the trigger can use it to pull the dog away from the real door. Here the defender's tool has a rule. The attacker is trying to turn that rule into an off switch.
Two cautions apply. ESET says the group was "apparently attempting" this. The report does not say whether it worked. It is also one group in one report, so it is a single data point, not a trend. Still, it offers a practical lesson. A security tool that refuses to analyse a file has not found the code safe. It has simply not looked.
What else the period showed
The AI findings sit inside a wider picture. Several items bear on the people who run companies.
ESET describes a shift at DeceptiveDevelopment, a group it links to North Korea. The group used to send fake job-interview lures. It has now moved to wider attacks on software supply chains. It steals developer credentials. It then uses them to slip malicious code into public and private code repositories.
Lazarus, another North Korea-linked group, is again going after the drone sector. ESET says it usually gets in through social engineering built around recruitment.
In Ukraine, ESET found a large campaign against MDaemon mail servers. The attackers installed two families of backdoor malware, named HELLAGE and HELLTAG. A backdoor is hidden software that lets attackers return later.
ESET also says FrostyNeighbor, a group it links to Belarus, hit a Ukrainian manufacturer that matters to the country's transport sector. ESET calls it the first destructive attack by a Belarus-linked actor that it has documented in Europe.
One limit applies to all of this. ESET says the report covers selected groups and only a fraction of the intelligence in its paid reports. It rests mostly on ESET's own telemetry. It is a view from one vendor's sensors, not a census.
Questions to put to your team
First, ask what your AI-assisted security tools do when they refuse or fail to analyse a file. A refusal should reach a human as an alert. It should not pass as a clean result.
Second, ask who can write to your code repositories. Ask what protects the developer credentials that grant that access. ESET's account of DeceptiveDevelopment shows stolen credentials being used to plant code.
Third, ask how engineers handle unsolicited recruiter contact. This matters most in technology, crypto and defense-related work. ESET describes recruitment-themed lures from two North Korea-linked groups.
Fourth, ask who owns patching of mail servers. Ask whether that owner knows about the MDaemon campaign ESET describes.
ESET's evidence suggests attackers use AI as a faster pen. One group has also tried to turn the defender's AI against its owner. Both need an answer from your team.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: ESET.
