AI in Cybersecurity: The Digital Arms Race

The same technology that protects our systems is being used to break them, and AI has kicked off an arms race in cybersecurity where both sides are running on the same engine. So who actually benefits more, the attackers or the defenders?

Official voice

In December 2025 the Swiss Federal Council published its assessment of AI’s impact on cybersecurity in response to Postulate 23.3861. Their conclusion: “AI does not fundamentally change cybersecurity, but acts as a catalyst for existing developments.”

A catalyst, not a revolution: what was already dangerous just got faster and harder to spot.

BACS Director Florian Schütz put it this way: “I have hope that AI will ultimately help defenders more than attackers.” Hope, not certainty, and when the head of Switzerland’s federal cybersecurity agency frames it like that you pay attention…

AI in defence

Security Operations Centres that used to spend 30 minutes or more investigating an alert can now do it in under two, which is not a small improvement but a completley different category of response.

The Darktrace 2026 report shows that 77% of security stacks now include generative AI, and about three quarters of organisations are already deploying agentic AI, meaning systems that do not just flag problems but act on them by themselves.

The big vendors are investing heavily:

  • Microsoft Security Copilot launched 12 new AI agents for threat triage, identity management and data security.
  • SentinelOne Purple AI Athena works as a fully agentic security analyst, investigating and responding without human input.
  • Palo Alto Networks paired its XSIAM platform with a large Google Cloud deal, combining AI-native security operations with hyperscale infrastructure.

The financial argument is also clear: organisations using AI in their security programmes save ca. $2 million per breach on average and contain incidents 80 days faster. Considering what a single breach costs these days, AI basically pays for itself (and then some).

AI on offence

And here is where it gets uncomfortable.

AI-powered phishing volumes surged by about 1,200 percent, vishing (voice phishing) jumped 170% in a single quarter, and AI-crafted phishing emails achieve four times the click rate of human-written ones. When researchers tested whether people could detect AI-generated deepfake voices the success rate was roughly a quarter, meaning three out of four deepfake calls work.

On the vulnerability side, research from the Cloud Security Alliance found that GPT-4 could exploit 87% of real one-day vulnerabilities when given their descriptions, and in January 2026 an AI agent found all 12 zero-day vulnerabilities in a new OpenSSL release.

The window between disclosure and exploitation has collapsed: in 2018 the average zero-day exploitation timeline was 756 days, but by 2024 researchers demonstrated exploitation in as little as four hours. The human patch cycle was not designed for this.

Over 200 malicious large language models now circulate on dark web marketplaces, built for crafting malware and generating social engineering scripts (GhostGPT for example is available for $50 per week, a bargain and cheaper than the defensive software). About three quarters of malware is now polymorphic, mutating every 15 seconds to avoid signature-based detection.

Switzerland is not immune

The BACS recorded 65,000 cyber incidents in 2025. Deepfake-driven investment fraud saw a fivefold increase with 5,706 reports filed, and total fraud losses reached about CHF 250 million.

In January 2026 an entrepreneur in Canton Schwyz lost several million francs after criminals cloned a business contact’s voice with AI. Several million from one phone call, and the technology to do this costs less than a dinner…

Fabricated videos of Federal Councillor Karin Keller-Sutter also circulated online, good enough to require expert analysis to debunk. If deepfakes can imitate the finance minister, how is a regular employee supposed to recognise a fake CEO call?

FINMA reports that half of supervised financial institutions already use AI, and President Marlene Amstad warned that this makes banks more efficient but also more vulnerable at the same time.

Switzerland is also producing answers though: Lakera from Zurich (AI security) was bought by Check Point for around $300 million, Invariant Labs (an ETH spin-off) went to Snyk, and Aurigin is working on deepfake voice detection. The talent and research are there; what is needed is a coherent strategy.

According to the World Economic Forum’s Global Cybersecurity Outlook 2026, 87% of organisations have already experienced AI-powered attacks. There are 4.8 million unfilled cybersecurity positions globally, and Gartner predicts that by 2027 over 40% of cybersecurity spending will be tied to AI, not because organisations want to spend more but because there is no alternative.

What to do

Start with your people, because security awareness training designed around text-based phishing emails is outdated when attackers deploy AI-generated voice clones and video deepfakes. Train staff to verify through secondary channels: every unusual request should trigger a callback on a known number, regardless of how familiar the voice sounds.

Then look at your patch management, because if vulnerabilities get exploited within hours rather than months, quarterly update cycles are a thing of the past, so automate where possible.

AI-powered defensive tools are worth adopting, but not every product labeling itself “AI-driven” delivers real protection. Look for measurable improvements in detection rates and response times: SOC automation, behavioural analytics, AI-assisted threat hunting are the areas where the returns are already proven.

For Swiss organisations specifically: implement voice verification protocols for financial transactions above a defined treshold, because multi-person authorisation for significant transfers is not optional anymore. And if your finance team has not been briefed on AI voice cloning yet, have that conversation this week.

Treat cybersecurity as a board-level concern, because the CHF 250 million in fraud losses reported to BACS in 2025 is only what was reported, and the real number is almost certainly higher…

We help organisations understand the AI-driven threat landscape and build defences to match. Get in touch.

Sources

Autor: Andriu Isenring CISSP