AI-powered cyberattacks: what really changes for your defence

Artificial intelligence did not give attackers brand new crimes. It made the old ones cheaper, faster and much more convincing. A phishing campaign that used to take a team days to localise can now be generated in minutes, in flawless French, Arabic or Tunisian dialect, and tailored to each recipient. For security teams, that changes the economics of defence more than the nature of the threat.

Phishing without typos

For years, awareness training taught people to spot clumsy wording and odd formatting. Large language models erased those signals. Attackers now combine public data (LinkedIn profiles, press releases, organisation charts) with generated text to write messages that sound exactly like a supplier, a manager or the bank. Criminal tools such as WormGPT and FraudGPT, advertised on underground forums since 2023, package this capability for people with no technical skill at all.

The human element remains the main entry point: the Verizon 2025 Data Breach Investigations Report finds a human action, such as clicking a link or falling for a pretext, in roughly 60% of breaches.

Deepfake executives and cloned voices

In early 2024, an employee of the engineering firm Arup in Hong Kong transferred about US$25 million after a video call in which the “CFO” and several colleagues were all deepfakes. Voice cloning is even easier: a few seconds of audio from a webinar or a voicemail greeting are enough to imitate someone convincingly on the phone.

The lesson is procedural, not technical. Any request to move money, change bank details or share credentials must be confirmed through a second, independent channel, no matter how familiar the face or the voice.

Faster reconnaissance and intrusion

AI also speeds up the technical side of an attack: sorting leaked data, mapping exposed services, explaining a vulnerability and drafting exploit code. In November 2025, Anthropic reported disrupting an espionage campaign, attributed to a state-sponsored group, in which an AI agent carried out most of the intrusion steps on its own, with human operators stepping in only at a few decision points. Vulnerability exploitation is already the initial access vector in about 20% of breaches according to Verizon, and automation shortens the time between disclosure and attack.

Your own AI is now an attack surface

Chatbots, copilots and AI agents connected to internal data create new risks. Prompt injection, ranked first in the OWASP Top 10 for LLM applications, lets an attacker hide instructions in an email, a web page or a document that the assistant will obey. Employees pasting sensitive data into unapproved tools (shadow AI) do the rest. IBM’s 2025 report found that 13% of organisations reported breaches of AI models or applications, and that 97% of them lacked proper AI access controls.

Defenders have AI too

The good news: the same technology works for the defence, and the numbers are measurable. According to IBM, organisations that use security AI and automation extensively cut the average cost of a breach by about US$1.9 million and shortened the breach lifecycle by about 80 days. The gain comes from triage and correlation: machines sort the noise so analysts can focus on the alerts that matter.

Seven controls that still work

• Phishing-resistant MFA (FIDO2 keys or passkeys) for email, VPN, cloud consoles and privileged accounts.
• Out-of-band verification for every payment, bank detail change or urgent request from an executive.
• Awareness training with realistic scenarios: AI-written phishing and deepfake calls, in the languages your teams actually use.
• Behavioural detection (EDR and NDR) that flags what a user or a machine does, not just known signatures.
• Automated SOC triage so that every alert is enriched and scored in seconds, around the clock.
• Risk-based patching of exposed systems, starting with known exploited vulnerabilities.
• An AI usage policy with approved tools, data rules and access controls for every model and agent.