How AI Is Changing the Cybersecurity Threat Landscape in Healthcare

Google’s Threat Intelligence Group findings show AI appearing at nearly every stage of the attack lifecycle, making existing tactics faster, cheaper, and harder to spot.

Reconnaissance is one of the clearest examples. Attackers are using AI to rapidly pull together open-source intelligence on a target, mapping organisational hierarchies, identifying decision-makers, and even researching salary bands for specific technical roles, work that used to take a skilled analyst hours or days, now compressed dramatically.

Phishing has changed in a similar way. AI-generated lures no longer carry the awkward phrasing or cultural mismatches that used to be reliable warning signs. Attackers are using models to write hyper-personalised messages that mirror an organisation’s professional tone, and in some cases to sustain believable, multi-turn conversations with a victim to build trust before ever delivering a malicious payload.

AI is also showing up inside the malware itself. One recently identified malware family calls an AI model’s API in real time to generate the code for its next stage of attack, then compiles and runs that code directly in memory so nothing is ever written to disk, a method designed specifically to dodge traditional file-based detection. Separately, researchers found a phishing kit impersonating a major cryptocurrency exchange that showed clear signs of being built using AI app-generation tools, lowering the skill and effort needed to produce a convincing fake.

Even attackers trying to build their own “custom” attack AI tools are mostly repackaging mainstream models. One underground toolkit marketed as a purpose-built offensive AI turned out, on investigation, to simply be commercial models stitched together with open-source tooling, suggesting that building dedicated criminal AI from scratch remains harder than buying access to legitimate tools and working around their guardrails. In one particularly novel case, attackers exploited the public sharing feature built into several mainstream AI chat tools, hosting fake troubleshooting instructions that tricked victims into pasting a malicious command straight into their own terminal.

It’s worth keeping this in perspective. Even the researchers tracking this activity closely note that they have not yet seen attackers achieve breakthrough capabilities that fundamentally change the threat landscape. What has changed so far is the speed, scale, and polish with which the attack has been executed. But for a sector like healthcare, where defenders are already stretched thin, and systems are often old and hard to patch, even modest gains in attacker speed and believability matter.