The end of luck: autonomous AI removes the last excuses for neglecting cybersecurity

Many organisations have avoided compromise so far through luck, not protection. Autonomous artificial intelligence takes luck out of the equation.
There is an old saying in cybersecurity: “There are companies that have been breached and companies that will be.” It was only ever half true. Between the two groups sat a third, very large one: vulnerable organisations that attackers simply never got around to.
That third group is disappearing. The criminals’ goal has not changed: money, data, access. What has changed is the cost of reaching it.
Why so many organisations have been spared until now
Attackers have always been limited by time, people and attention. Those limits worked as accidental protection for thousands of organisations that had invested nothing in security.
Some infrastructures were attacked deliberately; others fell by chance, after mass scanning. But theft was followed by work that few attackers ever finished. Data had to be sorted, credentials checked, access explored and its value assessed.
In practice it looked like this:
- Each group had its preferences: some went after particular industries, others after particular regions. The rest of the stolen data went unused.
- Credentials were checked by hand or with simple scripts. In manual checks, characters were mistyped or left out, and a perfectly valid account was filed under “not working”.
- Well-known admin pages were tried. If the developer had moved the login to an unusual address, the attacker gave up and moved on to the next target.
- Lists of RDP, VPN and other platform accounts were sold and resold, but a large share of them was never tried at all.
That is where the classic excuses came from: “we are too small”, “we have nothing of interest”, “nothing has happened to us so far”. Statistically, they even held up. Not because the organisation was protected, but because the attacker ran out of energy before reaching it.
From manual attacks to autonomous attacks
The difference between automation and autonomy is the difference between a script that gives up and an agent that keeps looking. Automation has been around for a long time and has made attacks faster, but it has stayed rigid: it does exactly what it was written to do and nothing more.
| Stage | Who does the work | What happens at the first obstacle | The limit |
|---|---|---|---|
| Manual | The attacker, step by step | Makes mistakes, gets tired, gives up | One person’s time and patience |
| Automated | Scripts and scanning tools | If the target does not match the template, moves on | The rigidity of the script |
| Autonomous | AI agents given a goal | Looks for another way in, tries variants, comes back | Computing cost, which keeps falling |
An AI agent is not given a list of steps but an objective. If the admin page is not at the usual address, it looks for it. If a password fails, it tries plausible variants. If one method is blocked, it picks another.
It works around the clock, on thousands of targets in parallel, with no preference for industry or geography. And whoever starts it no longer needs to be a specialist: knowing what they want to obtain is enough.
What an AI agent does from end to end
Once the goal is defined, time and resources are no longer a problem for the attacker. The agent works through the whole chain on its own, and the human steps in only for the important decisions:
- Collects. Gathers data from old and new leaks, from infostealer logs, from public sources and from scans.
- Correlates and sorts. Ties an account to a person, the person to an organisation, the organisation to its internet-facing systems.
- Prioritises. Ranks targets by value and by ease of access.
- Validates access. Tests credentials on RDP, VPN, email, admin panels and cloud services, including password variants.
- Performs reconnaissance. Maps the network and identifies systems, versions and vulnerabilities.
- Exploits and adapts. Chooses the right technique and changes approach when blocked.
- Expands. Gains higher privileges, moves laterally and secures a way back in.
- Extracts and analyses data. Decides for itself what is valuable: financial data, personal data, contracts, correspondence.
- Reports. Delivers a summary: where it got in, what it found, what can be exploited and how.
The attacker, experienced hacker or not, collects the results. From there on they only decide how to use them: extortion, fraud, selling the access or encrypting the data.
This is not a future scenario
Attacks run mostly by AI were already documented publicly in 2025. Two reports published by Anthropic, one of the developers of AI models, describe cases in which its own model was misused.
- August 2025, data extortion. A single actor used an AI coding agent for a data theft and extortion campaign against at least 17 organisations, including healthcare institutions, emergency services and government bodies. The AI carried out reconnaissance, harvested credentials, chose which data to take and wrote the extortion messages. Some demands exceeded 500,000 US dollars.
- November 2025, cyber espionage. A group attributed to a state actor targeted roughly 30 organisations in several countries. According to the report, the AI executed 80–90% of the operations, and humans stepped in at only 4–6 decision points per campaign.
The same report also notes the limits: the agent sometimes reported credentials that did not work, or presented public information as a discovery. AI agents make mistakes. The difference is that their mistakes cost almost nothing: they simply try again, endlessly. A single mistake by the defender can cost the organisation.
Two conclusions follow from these cases. One person with an AI agent does the work of a whole team. And what was within reach of a state actor in 2025 comes within everyone’s reach as the models get cheaper.
Stolen credentials: some expire, many do not
Part of all stolen data loses its value over time, but nobody can know in advance which part. Passwords change, especially where a password policy is actually enforced. An account stolen three years ago may be long useless.
Practice shows the other side too. Some credentials and accounts stay valid for years:
- service and technical accounts, exempt from password expiry “so the application does not stop”;
- accounts of former employees or former suppliers, left active;
- local administrator accounts and accounts shared by several people;
- VPN or RDP access set up “temporarily” and forgotten;
- API keys and tokens, which fall outside the password policy;
- passwords changed only in form: the same root with a different year or digit at the end.
Until now, few attackers had the patience to go back through old archives and separate dead accounts from live ones. An AI agent has that patience. It can systematically check everything that has ever been stolen and can infer the current password from the old one, starting from the pattern the person used to change it.
Personal data does not expire at all. Names, national ID numbers, addresses, job titles and professional relationships stay the same and feed phishing messages that are flawlessly written, in the victim’s own language and tone.
A password policy helps, but it is not enough. What matters is an inventory of accounts, disabling the unused ones, and multi-factor authentication, which makes a stolen password insufficient on its own.
What we have documented over the years
The theft of accounts from individuals, companies and public institutions in Romania is nothing new: we have been analysing and reporting it since 2020. The cases below show how much raw material already exists for an AI agent.
In June 2021, the Romanian news programme Știrile ProTV reported on a list put up for sale on the dark web: about 11 million accounts from 16 countries, including around 50,000 belonging to Romanians and about 2,000 compromised .ro domains. The accounts were personal, professional and institutional. Nobody knew when the data had been extracted; the theft came to light only when the list was offered for sale. We explained at the time, in the ProTV report (in Romanian), that the victims had been infected with data-stealing applications and that passwords had to be changed on every platform in use.
Lists like this do not disappear. They circulate, get resold and pile up in ever larger archives, and the other cases we have analysed show where they come from and how they are used:
- VPN accounts of institutions and companies (2021). In September 2021, a list of about 500,000 Fortinet VPN accounts was published on a hacking forum, with usernames and passwords in clear text, sorted by country and IP address. Those affected in Romania included public institutions, medical equipment distributors, travel agencies, telecom companies and a pharmacy chain. ProDefence had sent the list of vulnerable Romanian servers to CERT-RO as early as March 2021. Details in 500,000 Fortinet VPN accounts exposed: the role of early warnings (in Romanian).
- Email accounts of a ministry (2022). A fake page imitated the email server of Romania’s Ministry of Foreign Affairs, and the link was personalised with the address of an embassy. The goal was not financial gain but access to the institution’s infrastructure. Analysis: The Ministry of Foreign Affairs, target of a cyberattack (in Romanian).
- Passwords stolen from the browser (2020). Data-stealing malware collects browser-saved passwords first. We got into the control panel of one such program because the attacker himself had protected it with a weak password, guessed after a few attempts: Borr Malware: an attacker’s control panel and Hacking the Hacker (both in Romanian).
- Individuals’ identity and email accounts (2020). A phishing campaign imitating PayPal asked not only for card details but also for the email password and copies of identity documents. With those, the attacker could recover the victim’s accounts even after the passwords were changed: Data theft through PayPal phishing (in Romanian).
- From stolen credentials to unauthorised payments (2024). At a travel agency, stealer malware harvested the login credentials for the platforms in use, and the attackers entered the internal systems and made payments. The case was presented at the Romanian Tourism Fair: how stealer malware led to fraud at a travel agency.
- Bypassing two-step authentication (2022, 2024). Campaigns against customers of Romanian banks showed that codes received by SMS can be intercepted too: Romanian banks’ customers, targets of cybercriminals and Bank phishing and 2FA bypass.
The Fortinet case shows best what has changed. In 2021, six months passed between the warning in March and the publication of the list in September, and the organisations that acted in time had no cause for alarm. Checking 500,000 accounts took time and people back then. An AI agent works through such a list without a break, from the first account to the last, and does not skip any country.
We have also written about how AI makes attacks faster and more credible in Artificial intelligence and cyberattacks (in Romanian). Autonomy is the next step.
The excuses that no longer hold
Every reason given for postponing security measures rested on the same assumption: the attacker has no time for us. That assumption is no longer valid.
| The excuse | Why it no longer works |
|---|---|
| “We are too small to interest anyone.” | An AI agent does not pick targets by size. When the cost per target approaches zero, every target is worth trying. |
| “We have no valuable data.” | You have bank accounts, invoices, customer and employee data, access to your partners’ systems. The attacker decides what is valuable. |
| “Nothing has happened to us so far.” | The absence of a known incident does not prove the absence of access. Without monitoring, you have no way of knowing. |
| “We moved the admin page, nobody will find it.” | Hiding worked as long as the attacker gave up quickly. An AI agent searches until it finds. |
| “Those passwords are old.” | Some are. Service accounts, forgotten accounts and passwords changed to the same pattern are not. |
| “We have antivirus and a firewall.” | An attacker who logs in with a valid account triggers no alarm. They authenticate; they do not break anything. |
| “It costs too much.” | The cost of an attack has fallen dramatically. The cost of an incident has not: downtime, recovery, penalties, lost customers. |
| “We will deal with it after the audit.” | The agent does not check whether your evidence file is in order. It checks whether the measure works in reality. |
What remains to be done: the measures are known
Defence does not call for exotic technology, but for the consistent application of measures that have been known for years. AI agents have not invented new attack techniques. They exploit the same weaknesses, only faster and without tiring.
The practical priorities:
- Multi-factor authentication on everything exposed to the internet: VPN, RDP, email, admin panels, cloud services.
- An inventory of accounts and access: disabling unused accounts, rotating service account passwords, eliminating shared accounts.
- A smaller exposed surface: RDP and admin interfaces have no business being directly on the internet.
- Updates applied on time, especially on devices at the network edge.
- Monitoring and logs: a valid but unusual login has to be noticed by someone.
- Checking for leaked data: find out which of the organisation’s accounts are already circulating and change them.
- Tested backups and a continuity plan.
- An incident response plan, rehearsed before the incident.
- Staff training, because AI-generated phishing no longer contains language mistakes.
- Supplier security: their access to your systems is your risk too.
- Regular testing: a penetration test shows what the agent would find, before someone else sends it.
The list overlaps almost point by point with the risk-management measures in Article 21 of Directive (EU) 2022/2555 (NIS2), transposed in Romania by Government Emergency Ordinance no. 155/2024, approved by Law no. 124/2025. The Directive requires risk analysis, incident handling, business continuity, supply chain security, access control, training and multi-factor authentication.
NIS2 is often seen as a bureaucratic obligation. In reality, it describes the minimum level of protection any organisation needs against an attacker that no longer gets tired. And responsibility for approving and overseeing the measures lies with the management body, not with the IT department.
The measures are useful to organisations outside the scope of NIS2 as well. The AI agent does not check the legal classification of its target.
Defenders have AI at their disposal too, for detection, alert correlation and response. But the tools help only where the basics are already in place: it is known which systems exist, who has access to them and what normal behaviour looks like.
The end of an era
The era in which missing security measures could be offset by luck is over. The saying needs rewriting: there are organisations that have been breached and organisations whose data is waiting in the queue to be checked.
The useful question for management is no longer “who would attack us?”. It is “what does an agent find when it systematically checks, day and night, everything we have exposed and everything that has ever been stolen from us?”. The answer can be found before the incident, not after.
ProDefence, a NIS auditor accredited by Romania’s National Cyber Security Directorate (DNSC), works with organisations at any point along this road:
- Scoping: find out whether the organisation falls under NIS2 with the tool at cysnis.ro/incadrare (in Romanian).
- Legislation and obligations: an overview of the applicable framework, on the page NIS2 Directive in Romania.
- Consultancy and implementation: gap analysis, security measures, documentation and audit preparation, through our NIS2 consultancy service.
- Outsourced NIS2 officer, for organisations that do not have this competence in-house: service details.
- Penetration testing: checking what is exposed, before someone else does.
- Training for staff and management, including through the CysEdu platform.
Write to us using the form on prodefence.ro for an online discussion about your organisation’s situation.
Sources
- Anthropic. (27 August 2025). Detecting and countering misuse of AI: August 2025.
- Anthropic. (13 November 2025). Disrupting the first reported AI-orchestrated cyber espionage campaign.
- Știrile ProTV. (25 June 2021). Datele a 50.000 de români au fost furate și vândute de hackeri (in Romanian).
- Directive (EU) 2022/2555 (NIS2), Articles 20 and 21.
- Romanian Government Emergency Ordinance no. 155/2024 on the cybersecurity framework for networks and information systems in the national civilian cyberspace, approved by Law no. 124/2025.

