Anyone can build a case management system now. Then you have to keep it safe.

It has never been easier to build software. Describe a case management system to an AI on a Friday afternoon and by Monday you will have something that looks like one: a login page, care plans, incident forms, a dashboard with a pie chart nobody asked for. It will demo well.

Then you own it. Not for the length of the project, but for as long as anyone's records are in it. And the people trying to get in have the same tools you used to build it, with a great deal more practice.

The attackers got an upgrade

In April, Anthropic reported that its Claude Mythos Preview model had "autonomously identified and fully exploited a 17-year-old remote code execution flaw" in FreeBSD, "granting unauthenticated root access". Nobody set out to build that. Anthropic says these capabilities "emerged as a downstream consequence of general improvements in code, reasoning, and autonomy".

The following month, Google's Threat Intelligence Group confirmed the other half of the story: "For the first time, GTIG has identified a threat actor using a zero-day exploit that we believe was developed with AI." Anthropic had already written up a campaign in which the attacker used AI to perform 80 to 90% of the work.

Ollie Whitehouse, the NCSC's Chief Technology Officer, has told UK organisations to prepare for a "forced correction" of technical debt, because AI is "showing the ability to exploit this technical debt at scale and at pace". VulnCheck's figures put 28.3% of exploited vulnerabilities being used within 24 hours of disclosure, and Mandiant's M-Trends 2026 report found that exploits now routinely arrive before the patch does. The gap between a patch being available and someone being inside your system is shorter than most providers' change-control cycle.

The surface area is bigger than the code you wrote

Your system is never just your system. It is the framework, the database, the hosting, the two hundred open-source packages your developer pulled in without reading, the family app, the rostering integration, and forty support workers' phones. Malicious packages discovered on public repositories rose 75% in Sonatype's latest count. Each of those is a door, and you are now responsible for all of them.

Then there is the code itself. Georgia Tech researchers scanning 43,000 security advisories found 74 confirmed vulnerabilities introduced by AI coding tools, 14 of them critical. Their advice: "If you're shipping AI output to production, review it the way you'd review a junior developer's pull request." That assumes you have a senior developer to do the reviewing.

Care is not a quiet corner of the internet

If you are thinking "we're a small provider, who would bother?", the sensors disagree. SonicWall recorded 264,000 intrusion events against UK healthcare clients in the first five months of 2026, against 27,000 for the whole of 2025. The top attack vector was Log4j, a vulnerability from 2021, which tells you how much unpatched software is still running in the sector.

HCRG Care Group, which provides health and social care for the NHS and councils, had 2.275 terabytes taken and a $2m ransom demanded. Nearly two years after the Synnovis attack, one London trust was still "relying on paper processes and manual uploads". Care records are among the most sensitive data anyone holds: safeguarding notes, medication, the histories of children in care. There is no reason to think they will be treated as an exception.

What "having a security team" actually means

Not the person who fills in the DSPT once a year. It means someone whose job, every week, is patching within days rather than quarters. It means knowing exactly what you have exposed to the internet and keeping that list as short as possible, which is the NCSC's first recommendation. It means monitoring that notices at 3am, an incident plan that has been rehearsed rather than filed, and updating by default rather than when someone gets round to it.

That is a standing cost, not a project with an end date. If your build plan does not have a line for it, you have not costed the build.

Where we stand

We build this for a living, so I am not neutral. The honest version is that the security bill is the part of building your own system that does not appear in the weekend demo, and it never stops arriving. Whether you buy from us, from someone else, or build in-house, ask the same questions. Who patches, and how fast? What is exposed to the internet, and how do you know? Who is watching at 3am? What happened the last time something broke?

If the answer to any of those is a shrug, the dashboard is not going to help.