- Proofpoint and Bitdefender researchers told Dark Reading that AI has made some attackers noisier and less skilled, while the most careful groups avoid it.
- A widely repeated figure says 84% of ransomware attacks use AI. It comes from self-reported survey data and does not show that attacks got better.
- Ask what any AI-threat statistic measures, and check that your team can handle more low-quality volume while still blocking the plain, common attacks.
Picture a board paper that says 84% of ransomware attacks now use AI. The obvious conclusion is that attackers have become much stronger. New reporting from Dark Reading suggests that conclusion deserves a second look.
The lesson here is that AI multiplies output, and the multiplier applies to mistakes too. For some attackers, more automation has meant more errors. It has also meant more of their own infrastructure left in plain view.
What the researchers reported
Dark Reading noted that mainstream media this week repeated a figure from self-reported survey data. It says about 84% of ransomware attacks use AI in some form. That tells you AI is present. It does not tell you the attacks improved.
Proofpoint offered a different view in a briefing shared with Dark Reading. Selena Larson, its principal threat researcher, rejects the idea that AI is making attackers better. "It's just not the case," she said.
How automation hurt one criminal group
Proofpoint tracks a group it calls TA4903. The group has run credential phishing and business email compromise (BEC) for five years. BEC means criminals pose as a trusted sender to steal money or logins.
By Larson's account, the group was once effective. It posed as the US government. It used detailed PDFs, well-made QR codes and convincing landing pages.
In March, Proofpoint says, the group let AI handle each stage of its attacks. That covered lures, targeting and password theft.
In the last six months, it also began "vibe coding". That means asking an AI to write software and using whatever comes back. The group tried to build its own phishing kits this way.
The kits were left exposed, Proofpoint says. Researchers could see control panels, Cloudflare deployments, tools for creating links and lures, and stolen credentials.
Larson said this lets researchers see where the group's infrastructure sits and who it may target next.
The phishing messages also got worse. Some QR codes did not display in email. Some emails carried an elaborate PDF but no text. Others posed as real organizations but still held the default text the AI had produced.
State-backed groups show the same pattern
Researchers at Genians tied a campaign against Ukraine to Kimsuky, a North Korean group. They said Kimsuky used AI agents and large language model infrastructure. The AI wrote phishing content and code.
The results were messy. The emails confused their targets, mixing Ukrainian, South Korean and Japanese recipients. The writing swung between formal and casual. The text did not match the decoy documents. Google Meet invites held Zoom links. The malware had logic errors and missing parts.
Bitdefender reported in March that Pakistan's APT36 was transitioning from off-the-shelf malware to "vibeware". These are throwaway files generated on demand. The volume helped, because removing one or two implants made little difference.
Dark Reading adds that some of the output failed. One stealer could not send data out. A bot management system did not work as intended.
Why careful attackers hold back
Martin Zugec, technical solutions director at Bitdefender, gave a practical reason. Espionage depends on staying quiet, and AI does not. Given a task, it tries the direct route and then tries others. That creates noise.
He also said that when researchers document an operation, they can set it back by years. So a serious group has little reason to add a tool that raises its chance of being found.
AI is not absent from serious campaigns. In August, Bitdefender described SilkParasite, a China-linked group. It used AI in a narrow role, helping to develop malware. That is limited use, not broad adoption.
Two caveats apply. This is one article built on researchers' views of the groups they track. And Zugec himself calls the publicly documented AI threats "anecdotes".
What leaders should do with this
Zugec says most of what researchers still see is "very boring" breaches. Examples are ClickFix and living-off-the-land attacks. In the second type, criminals use tools already on the victim's systems. Plain attacks still work, so basic controls still matter most.
Five questions to put to your security team and vendors:
First, when a vendor quotes an AI-threat statistic, ask what it measures. Does it show AI was present, or that attacks got better? Who was surveyed, and how?
Second, ask whether your defences are tested against ClickFix and living-off-the-land techniques. AI-themed scenarios alone are not enough.
Third, ask whether analysts can handle a rising volume of poor phishing without missing the one good message.
Fourth, ask whether your threat intelligence tracks research on attackers' exposed infrastructure. That is where sloppy automation shows.
Fifth, check that spending on AI-specific defences matches what is actually hitting you.
A photocopier makes copies faster, typos included. Attackers who automate without review copy their mistakes at scale. The sloppy groups are the ones researchers can see.
It follows, in our view, that careful groups are harder to see. They add less automated clutter, so there is less for researchers to find. That is our inference, not a finding in the report.
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: Dark Reading.
