Μετάβαση στο περιεχόμενο

Το καλάθι σας

Το καλάθι σας είναι άδειο.

Προσθέστε μια υπηρεσία από το κατάστημα και θα σας περιμένει εδώ μέχρι την αγορά.

Δείτε το κατάστημα

Μετά την πληρωμή

  1. Η ομάδα διαβάζει τα στοιχεία σας και σας στέλνει email, συνήθως μέσα σε μία εργάσιμη ημέρα.
  2. Παρακολουθείτε τη δουλειά και μιλάτε με την ομάδα στην πύλη πελατών.
Μερικό σύνολο

Εφάπαξ χρεώσεις υπηρεσιών, σε USD.

Ταμείο Δείτε το καλάθι
Απειλές ΤΝ

AI-run cyberattacks are here: what changed, and why speed matters

Attackers now hand break-in work to AI agents. What the AI labs and threat teams have documented since 2025, and what it means for your business.

Δημοσίευση: 5 λεπτά ανάγνωσης8 πηγέςΤα άρθρα είναι στα αγγλικά.

For years, a serious break-in needed a team of skilled people working for weeks. That is no longer true. Attackers now hand most of the work to AI agents that scan, probe, write exploit code and steal data at machine speed, with people stepping in only at key moments.

This is not a forecast. The AI labs and threat intelligence teams have published the cases.

What has already happened

An espionage campaign run mostly by AI. In November 2025, Anthropic disclosed that a group it assessed with high confidence to be Chinese state-sponsored had used its Claude Code tool to attack roughly thirty organisations: large tech companies, financial institutions, chemical manufacturers and government agencies. A small number of the attacks succeeded.

The humans picked the targets, built the framework and tricked the AI past its safeguards by posing as defensive security testers. The AI did the rest: it mapped the target systems, found and tested vulnerabilities, wrote exploit code, harvested credentials, sorted the stolen data and wrote up its own notes for the next stage. Anthropic says the AI performed 80 to 90 percent of the campaign, with humans stepping in at perhaps four to six key decision points per campaign. At its peak it made thousands of requests, often several per second. No human team can work at that pace.

Extortion of 17 organisations by one person. In August 2025, Anthropic described a criminal who used the same kind of tool against at least 17 organisations, including healthcare, emergency services, government and religious institutions. The AI automated reconnaissance, credential theft and network access. It also chose which data to steal, studied the victims’ finances to set ransom demands, and wrote the ransom notes. Some demands went above $500,000.

Ransomware built by someone who could not build it. The same report describes a criminal who used AI to develop ransomware and sell it on forums for $400 to $1,200 a package. Anthropic says the actor appears to have depended on AI to build working malware.

Malware that asks an AI what to do next. In November 2025, Google’s Threat Intelligence Group reported malware that calls an AI model while it runs. One family, used by the Russian government-backed group APT28 against Ukraine, asks a model to generate the commands it then executes on the victim’s computer. Google called this “a new operational phase of AI abuse”.

AI agents moving through networks in minutes. In October 2026, CrowdStrike reported four previously unreported threat actors using AI agents in real intrusions, documented between May and August 2026, against healthcare, government, retail, technology and other targets. In one case, an AI agent moved across a network and harvested domain controller credentials in about 20 minutes. CrowdStrike says agentic AI working at machine speed can reach its goals “often within an hour”.

What changed

Three things.

  1. Skill is no longer the limit. Anthropic’s own conclusion: “The barriers to performing sophisticated cyberattacks have dropped substantially.” Groups with little experience and little money can now run attacks that once took a skilled team.
  2. Scale is cheap. An AI agent does not get tired. It keeps probing websites and accounts around the clock. Google also reports a maturing underground market selling AI tools for phishing, malware and vulnerability research.
  3. Speed has collapsed. The UK’s National Cyber Security Centre (NCSC) says “the time between disclosure and exploitation has shrunk to days and AI will almost certainly reduce this further.” Google’s Mandiant analysts measured the average time to exploit a new flaw falling from 63 days in 2018 to 2019 to 5 days in 2023, with outliers set aside. That was before AI agents joined in.

Why speed matters for you

When a flaw in a common plugin, framework or server is published, attackers start scanning for it. With AI, the scan, the test and the first attack can follow within hours. The NCSC expects AI to “almost certainly” increase attacks against systems that have not been updated.

So the question for a business is simple: would you find the weakness before an automated attacker does?

Most small businesses would not. The website was built once, the plugins were updated now and then, and nobody has looked at it from the outside since.

Not unstoppable

The same campaigns show the limits. Anthropic noted that the AI sometimes invented credentials or claimed to have found secrets that were already public. Attacks still need a way in: an outdated component, an exposed admin page, a weak sign-in, a misconfigured server.

Defenders use AI too. In July 2025, Google said its Big Sleep AI agent found a critical SQLite flaw that attackers were preparing to use, and that it was cut off before they could use it. Google called it the first time an AI agent had directly foiled an attempt to exploit a flaw in the wild.

The NCSC puts it plainly: “Keeping pace with ‘frontier AI’ capabilities will almost certainly be critical to cyber resilience for the decade to come.” It warns of a growing divide between systems that keep pace and systems that fall behind.

What to do now

  1. Know what you expose. List every website, subdomain, admin page and API that is reachable from the internet. Attackers’ tools find these in minutes.
  2. Patch fast. Turn on automatic updates where you can. For everything else, set a rule: critical security updates within days, not months.
  3. Remove what you do not use. Old plugins, test pages, forgotten subdomains and unused accounts are easy targets.
  4. Lock down sign-in. MFA on every admin account, and no shared passwords.
  5. Test from the outside, regularly. A one-off review from years ago says nothing about today’s threats.

Where EWC fits

AI attack methods change fast. Checklists and scanners written last year do not keep up. Our security team follows new AI attack methods and brings them into how we test.

Our Advanced black-box penetration test attacks your website from the outside, the way a real attacker would: reconnaissance, injection, file uploads, APIs and targeted exploit attempts. You get every finding with evidence, its severity and the fix, so you close the gaps before an automated attacker finds them.

Πηγές

Σε τι βασίζεται αυτό το άρθρο.

8 πηγές
  1. AnthropicDisrupting the first reported AI-orchestrated cyber espionage campaignanthropic.com
  2. AnthropicDetecting and countering misuse of AI: August 2025anthropic.com
  3. Google Threat Intelligence GroupAI Threat Tracker: Advances in Threat Actor Usage of AI Toolscloud.google.com
  4. CrowdStrikeCrowdStrike Uncovers New Agentic AI Campaignscrowdstrike.com
  5. CrowdStrikeAgentic AI attack campaignscrowdstrike.com
  6. NCSCImpact of AI on cyber threat from now to 2027ncsc.gov.uk
  7. Google Cloud (Mandiant)How Low Can You Go? An Analysis of 2023 Time-to-Exploit Trendscloud.google.com
  8. The RecordGoogle says 'Big Sleep' AI tool found bug hackers planned to usetherecord.media

Πού βοηθάμε · Επίπεδο 2

Προηγμένος έλεγχος διείσδυσης black-box

Ελέγχουμε την ιστοσελίδα σας απ’ έξω, όπως θα έκανε ένας πραγματικός επιτιθέμενος, και τεκμηριώνουμε κάθε εύρημα.

Τι περιλαμβάνει

  • Τεχνικές πραγματικού επιτιθέμενουΑναγνώριση, εγχύσεις, μεταφορτώσεις, API και απόπειρες εκμετάλλευσης.
  • Αυστηροί κανόνες εμπλοκήςΓραπτή εξουσιοδότηση, καμία αλλαγή ή διαγραφή, περιορισμένος ρυθμός αιτημάτων.
  • Πλήρης αναφορά ελέγχου διείσδυσηςΕυρήματα σε απλή γλώσσα, σοβαρότητα CVSS, διορθώσεις και επαλήθευση.