The Hype Machine
You search for the AI assistant everyone at work is talking about. A site comes up offering a “Pro” version. You download it, it installs, it works exactly as promised. Nothing looks wrong. (This scene is illustrative, but the trap it describes is real, and we will get to it.)
Meanwhile, somewhere else, the program you just installed is quietly opening a back door to your computer.
That is the AI threat most of us actually face. Not the movie version, where artificial intelligence hands criminals superpowers. Something duller, closer to home, and far more fixable.
The claim, and the report that deflates it
Every year the European Union’s cybersecurity agency, ENISA, publishes a large report on the threats facing Europe. The 2026 edition, published in September, analyses 8,257 incidents recorded during 2025 (ENISA, 2026, p. 9). Most of it is about familiar things: denial-of-service attacks, ransomware, phishing and fraud.
Threaded through it is a clear message about AI, and it is not the one the headlines usually carry. Here is my reading of it, as a proposition you can test against the evidence below: attackers are using ordinary AI tools to do old tricks faster, and the AI tools we invite into our own lives are becoming the target. The people who pay are the ones who trusted the download, the browser extension or the cloud account.
A speed boost, not a superpower
ENISA’s assessment is measured. Reporting over 2025, it says, “indicates that attackers primarily use consumer-grade AI tools to augment existing skills and adapt attack vectors rather than to achieve breakthrough capabilities” (ENISA, 2026, p. 15). For criminal and state-linked groups alike, the point was mainly to improve “the speed and success rates” of what they already did.
In practice, that was unglamorous. Attackers used large language models to write tailored phishing messages at scale, to produce material for influence campaigns, and to create fake identities for schemes in which people pose as remote workers (p. 15).
AI was also used for reconnaissance, writing scripts and tasks after a break-in, sometimes through “turnkey” AI-assisted toolkits sold on underground markets (p. 16). In fraud, ENISA says AI “further facilitates and increases the tempo” of fake websites built to lure people into scams (p. 42).
The underlying picture has not changed much. Among the social engineering techniques ENISA identified, phishing accounted for 77.8 percent (p. 11). Looking ahead, the agency still sees AI acting mainly as “a force multiplier,” although it expects more stages of attacks to be “directly enabled by AI” during 2026 (p. 96).
So the superhacker is, for now, mostly a very fast copywriter.
Receipt one: the “Pro” download
Back to that download. ENISA singles it out: “A fake Claude AI website, for example, was used to distribute malware” (ENISA, 2026, p. 17).
The source ENISA cites, a security analysis by Malwarebytes, describes a website whose domain mimicked the official one and offered a download billed as a “Pro” version of the assistant. The trap was well built: the app installed and ran as expected, so the victim saw nothing wrong.
In the background, it installed a remote access trojan called PlugX, a type of malware whose variants can log keystrokes and steal credentials (Dasic, 2026). The analysis does not say how many people downloaded it.
This is an old trick in new packaging. Criminals have long disguised malware as popular software. AI assistants are now among the most popular software there is, new versions appear constantly, and many people are not sure what the official download looks like. That combination is exactly what scammers look for.
Receipt two: eight minutes
The second example is about speed. ENISA reports that “an AI-assisted cloud intrusion reportedly achieved administrative access within 8 minutes” (ENISA, 2026, p. 17).
The underlying report, from the security firm Sysdig, describes an attack on an Amazon Web Services account in November 2025. The way in was not clever: valid credentials had been left in publicly accessible storage buckets, and those buckets held data used to feed AI models.
From there, Sysdig says, the attacker moved to administrator control, created a backdoor account, ran the victim’s own AI models at the victim’s expense and launched expensive computing resources (Sysdig, 2026). Follow the money and the victim paid twice: once in lost control, and again in the bill.
Two hedges are worth keeping. Sysdig infers AI involvement from clues in the attacker’s code and says the code suggests it was generated by an AI model. And the eight minutes is Sysdig’s figure for the core sequence of the break-in; its full timeline of the attack runs to about two hours.
The lesson is less about AI than about time. If an attacker can go from a leaked password to full control in minutes, a defence that relies on someone noticing and responding later is already too slow. The protection has to be in place beforehand: credentials kept out of public places, and accounts that cannot do more than they need to.
Key idea: Attack surface
Your attack surface is everything an attacker could try to get through: every app, account, password and connection. Each time you connect an AI assistant to your email, files or browser, you add to it, because anyone who takes over the assistant inherits whatever it can reach. The defence is called least privilege: give each tool only the access it needs for its job, and nothing more. A smaller attack surface means a smaller disaster when something does go wrong.
Receipt three: the extension reading your chats
This is the part of the report most relevant to ordinary users. “The integration of AI systems into enterprise environments creates a new attack surface,” ENISA writes (p. 15).
It goes further: “AI applications and their ecosystems are increasingly becoming targets, particularly where they have access to files, credentials, browser sessions or development environments,” and AI services “may also provide a path into trusted environments” (p. 17). Attackers are also experimenting with ways around AI models’ safeguards, including indirect prompt injection, where instructions are hidden in content the AI reads (p. 16).
One of the sources ENISA cites shows what this means in practice. In March 2026, Microsoft’s security researchers described malicious browser extensions for Chrome and Edge that impersonated legitimate AI assistant tools. They collected the websites people visited and the content of their conversations with ChatGPT and DeepSeek.
Microsoft cited outside reporting of roughly 900,000 installs and said its own data showed activity in more than 20,000 enterprise organisations (Microsoft Defender Security Research Team, 2026).
Think about what people type into AI chats: work documents, health worries, financial questions, drafts of difficult emails. An extension that quietly copies those conversations is a privacy breach of an unusually intimate kind. And the more we connect AI assistants to our inboxes, files and accounts, the more an attacker gains by compromising one.
The best case against me
Here is the strongest objection: if AI isn’t creating new kinds of attack, why worry about it at all? Isn’t this just ordinary cybercrime with a fashionable label?
Partly, yes. That is exactly what ENISA says. But speed and scale are not trivial. A phishing message that is better written, a fake site that goes up faster, a break-in measured in minutes: each one narrows the gap between a mistake and a disaster. And the new part is real, too. The tools themselves, sitting in our browsers and inboxes, are now worth stealing.
A footnote of our own: ENISA is open about its own use of AI. Its report states that AI was used “in a limited capacity” to help collect and process data, that every output was reviewed by experts, and that no AI-generated content appears in the report (p. 4). That kind of disclosure is worth copying.
The verdict
The evidence supports the duller story. The danger is less that AI makes attackers brilliant and more that we have handed AI tools the keys to our files, chats and accounts, often without checking where the tool came from. That is good news, in a way. Duller problems have duller fixes, and most of them are in your hands.
What we still don’t know
ENISA’s figures come mainly from open sources and information shared voluntarily, and the agency says plainly that these “do not constitute a complete picture” (ENISA, 2026, p. 9). The report does not estimate how much AI-assisted attacks cost Europe, and it does not claim AI has created new kinds of attack.
The examples above come from security companies’ own investigations, cited by ENISA; we read them, but we cannot independently confirm them. We do not know how many people installed the fake app. The eight-minute figure is described by ENISA as “reportedly” and by Sysdig as an AI-assisted attack it inferred from the attacker’s code, not one the attacker admitted.
Finally, this report covers incidents during 2025. ENISA’s separate assessment of the most advanced AI models is a different document, and it raises the question this one doesn’t answer: what happens when the superpower stops being hype? We will look at it separately.
What to do this week
- If you use AI apps: download them only from the official website or your device’s official app store, typed in yourself rather than reached through an ad, and be wary of any “Pro” or “unlocked” version offered elsewhere (Dasic, 2026).
- If you use browser extensions: look at the list today. Remove anything you do not recognise or no longer use, especially extensions that promise AI features and can read every page you visit.
- Before you connect an AI assistant to your email, files or calendar, ask whether it really needs that access, and how you would turn it off. Treat what you type into AI chats as something that could one day be read by someone else.
- If you run IT or security: list every AI tool, extension and integration in use, check what each can reach, keep credentials out of shared or public storage, and make sure the accounts AI systems use cannot do more than their job requires.
Go further. Operational AI Readiness is FairByDesign’s field manual for AI that is already live, and the complete Standard edition is free to read and download. Chapter 2, Where Risk Appears in Live AI Systems, maps the places AI quietly widens your exposure, and Chapter 7, Security Controls for LLM Apps and Agent Workflows, covers how to limit what an AI tool can reach and do. Read it free, or test one workflow with the free readiness assessment, paying close attention to domain 1, inventory and scope, and domain 7, LLM and agent security boundaries.
If you answer for AI at work, ENISA’s warning that AI tools are becoming targets lands on your desk, not just your security team’s. Mercury Intelligence ($10 a month) gives you our full investigations and a governance impact analysis of each one: what the finding means for your controls, the evidence a reviewer will expect, and the questions to put to your team this quarter. See what is included, or save with the annual plan.
FairByDesign is reader-funded. If this helped, get the free Mercury Brief or see ways to support. Support never buys editorial control.
References
Bakos, V. (2026). Operational AI readiness: A practitioner guide to governing AI that is already live. FairByDesign. https://fairbydesign.org/operational-ai-readiness/
Dasic, S. (2026, April 10). Fake Claude site installs malware that gives attackers access to your computer. Malwarebytes. https://www.malwarebytes.com/blog/scams/2026/04/fake-claude-site-installs-malware-that-gives-attackers-access-to-your-computer
European Union Agency for Cybersecurity. (2026). ENISA threat landscape 2026. ENISA. https://doi.org/10.2824/0806036. https://www.enisa.europa.eu/sites/default/files/2026-09/ENISA%20Threat%20Landscape%202026_Final.pdf
Microsoft Defender Security Research Team. (2026, March 5). Malicious AI assistant extensions harvest LLM chat histories. Microsoft Security Blog. https://www.microsoft.com/en-us/security/blog/2026/03/05/malicious-ai-assistant-extensions-harvest-llm-chat-histories/
Sysdig. (2026, February 3). AI-assisted cloud intrusion achieves admin access in 8 minutes. Sysdig Blog. https://www.sysdig.com/blog/ai-assisted-cloud-intrusion-achieves-admin-access-in-8-minutes



Leave a Reply
You must be logged in to post a comment.