Read the NCSC AI ransomware warning as a warning about malware and you will scope the wrong test. The assessment is about uplift: the same intrusions, cheaper, faster, and available to people who previously could not run them. The new malware is a forecast. The new exposure is already sitting on your network, holding a service token, wired into SharePoint, and it is the assistant your product team switched on four months ago.
This piece states what the warning says, marks the line between what is evidenced today and what is projected, then turns it into seven tests you can actually commission. The tests are deliberately mundane. Ransomware readiness work usually is.
What the NCSC AI ransomware warning actually says
The core judgement has been consistent since the NCSC published its assessment of the near-term impact of AI on the cyber threat in January 2024: AI will increase the volume and heighten the impact of cyber attacks, and the uplift is uneven. Reconnaissance and social engineering gain the most. Lateral movement, exploitation and exfiltration gain the least, at least for now, because they depend on environment-specific knowledge that models do not have.
The second effect is the one that matters for ransomware economics. AI lowers the barrier to entry. A low-capability actor who could not previously write a convincing lure in English, triage a victim list, or read a codebase quickly can now do all three. Capability that used to be scarce becomes rented.
The NCSC has since extended the line rather than softened it, judging it highly likely that AI-enabled malware will feature in a ransomware attack against UK infrastructure by 2028.
What “highly likely”means, precisely
It is not a figure of speech. UK assessment bodies use the Professional Head of Intelligence Assessment probability yardstick, where “highly likely”maps to roughly an 80 to 90 per cent probability band, sitting above “likely”and below “almost certain”. So the judgement is not “this might happen one day”. It is a near-term planning assumption with a date attached, written by people who are careful about which word they pick.
Evidenced now, forecast later
Keep these two columns apart when you brief a board, because conflating them is how you get asked for a budget you cannot justify twice.
- Evidenced today: phishing and business email compromise lures that no longer fail the grammar test, faster victim triage from open sources, quicker reading of leaked data and source code, faster vulnerability research against known bug classes.
- Forecast: adaptive payloads, autonomous intrusion from initial access through to encryption, malware that reasons about your environment in flight.
Initial access has not changed. It is still phishing, exposed services and stolen credentials. What changes is the hit rate and the clock. That is the whole point of an uplift assessment: the doors are the same, more people can open them, and they open faster.
What the warning changes about your attack surface
Here is the part the news coverage skips. If attacker effort gets cheaper, the first controls to fail are the ones that quietly assumed attacker effort was expensive. Rate limits tuned to human typing speed. Approval steps that work because someone has to notice the request. Anomaly thresholds calibrated against a volume of attempts that no longer holds. Go and look at those three before you look at anything branded.
The genuinely new surface is internal. Copilots wired to the document store. Retrieval systems with no tenancy boundary. MCP servers connecting a dozen tools to one orchestration layer. Coding agents with repository write access and shell execution.
Sit with that last one. An agent with repository write access, a build runner and a shell has the capability profile a ransomware operator normally spends a fortnight acquiring: code execution, persistence, credential reach, and a legitimate identity that your detection stack already trusts. Nobody had to deploy a dropper. Procurement did it.
Five factors for ranking what to test first
Ignore vendor risk scores. Judge each deployment on five things:
- Reachability: can untrusted text reach the model, from a supplier, a customer, a web page or a file?
- Capability: what tools can it call, and which of them write, delete, pay or deploy?
- Data breadth: what can it read in a single session, and does it respect the permissions of the person asking?
- Checkpoints: what genuinely requires a human approval, and is the approval legible enough to be meaningful?
- Reversibility and detectability: would you see the action in an audit log, and could you undo it?
Run that scoring honestly and you will often find a dull chatbot beats the flashy agent for risk: no tools, but it reads every file in the tenancy and logs nothing. The agent is behind approvals. The chatbot is a silent exfiltration channel. Our longer treatment of how to judge prompt injection exposure works through the same ranking in more detail.
Seven tests that turn the NCSC AI ransomware warning into evidence
1. AI-sharpened phishing, tested against the chain not the score
Stop measuring click rates. Measure the chain: time from first delivery to first report, whether the secure email gateway flagged anything, whether the SOC correlated two reports from different departments, whether the reported sample reached a human within the hour. Use lures written to the standard that is now trivially available, including correct internal terminology scraped from your own public material. Pass condition is a working detection and response chain, not an awareness percentage.
2. Indirect prompt injection through your real ingestion channels
Injection rarely arrives through the chat box. It arrives in a ticket description, a supplier PDF, a calendar invite, a fetched web page, or HTTP request and response headers that the agent never shows the user. Plant canary artefacts in each channel and see which ones the assistant obeys. Our walkthrough of injection carried in HTTP headers covers the channel most teams forget entirely. Test the ingestion path. The user interface is theatre.
3. Tool abuse and excessive agency
Once you have a working injection, the question is what it reaches. Enumerate the tool schemas the agent actually holds, not the ones in the design document, then prove which can be invoked without approval: write to a share, delete a record, raise a payment, push a commit, run a shell command. The blast radius tests for agent prompt injection set out the sequence. Remove the tools and it is a chatbot. Leave them in and it is an ai agent privilege escalation path with a trusted identity attached.
4. MCP server exposure
Three things to prove: that a poisoned tool description can steer the orchestration layer, that a second connected server can shadow or redefine a tool exposed by a trusted one, and that a tool can change its behaviour after the user has already approved it. Approval at install time is not approval at call time. If your client caches the first and never re-checks, that is the finding.
5. Coding agent supply chain
Plant an instruction in a repository README or an AGENTS.md file and see whether the agent follows it without the developer reading the diff. Check whether it installs packages it invented: attackers register hallucinated names on public registries precisely because agents keep asking for them, and a single install gives code execution inside your build pipeline. Then check whether a build step can post secrets outbound. We cover the file-level case in testing and containing AGENTS.md injection.
6. Credential and secret reach from AI service accounts
Take the token the assistant runs as. Map it. File shares, object storage, ticketing, source control, backup targets, and whether it inherits the permissions of the requesting user or a standing service identity. One internal assistant connected to SharePoint and a ticketing queue, driven by a supplier-submitted ticket body, can summarise and forward credential material to an external address without any malware involved. Data breadth proven, payload never written.
7. Backup and restore with an agent in the picture
Run the restore exercise you already run, then add one question: can any AI service identity reach backup storage, snapshot schedules or retention settings? Immutability that an agent identity can toggle is not immutability. Cyber Essentials-level hygiene and your existing tabletop cover the rest; this is the hour you add to them.
How to prove a finding rather than assert one
These systems are non-deterministic, which breaks normal reporting discipline. Run every proof of concept at least twice and record the reproduction rate. A path that fires two times in ten is still real, but it is a different remediation conversation from one that fires every time, and a finding that vanishes on re-run and goes in the report anyway is the fastest way to lose a board on this subject.
Separate model assertion from system behaviour. A chatbot saying it deleted the file is a measurement of text. It is not a measurement of harm. Go and check the file, the audit log, the captured request. Every fact in the report traces to one of those three or it does not go in.
Then collapse duplicates. One injection path reachable through eleven document types is one finding with eleven channels, not eleven findings, and inflating the count inflates the risk picture in a way an engineer will notice and resent. Our LLM penetration testing checklist works through this evidence standard properly. Proof, not output.
Mistakes made in the week after the headlines
Buying AI-branded detection before testing whether the AI systems you already run can be hijacked. The spend is visible, the exposure is not, and the order is backwards.
Scoping confusion. Testing an AI system as a target and using AI to assist a conventional test are different engagements with different rules of engagement, different authorisation needs and different evidence standards. Write which one you are buying into the first page of the scoping document.
Treating a model provider’s safety claims as a control you own. You do not own their filter, you cannot audit it, and it will not appear in your incident timeline.
Accepting assessment output with identifiers that do not exist, remediation steps that do not apply to your platform, or severity assigned with no demonstrated consequence. Those are the recurring red flags in model-assisted security reports, and they are worth a verification pass every time.
And excluding the assistant pilot because it is “only a trial”while it quietly reads the entire document store. Scope it. Pilots have tenancy too.
A sensible order of work for the next two quarters
Inventory first: every deployment with tool access, including the ones bought on a departmental card. Rank by blast radius, not by how impressive the demo looked. Fold the seven tests into the ransomware readiness exercise you already run rather than standing up a separate AI stream nobody owns. Then set re-test triggers in writing: a new tool, a new connector, a new model version, a permissions change. Any one of those invalidates the last result.
The Short Version
The ncsc ai ransomware warning is an uplift assessment with a dated probability attached, not the announcement of a new malware family. The controls that break first are the ones that assumed attacker effort was expensive. The most testable new exposure in your estate is the assistant holding live credentials, not the payload a criminal group might ship in 2028. Test reachability, capability and reach, run every proof twice, and report consequence-proven findings a developer will actually fix. If you want the ground work before commissioning, start with how agent tool use should be constrained.
Frequently Asked Questions
What does the NCSC AI ransomware warning actually say?
Since its January 2024 report on the near-term impact of AI on the cyber threat, the NCSC has assessed that AI will increase the volume and impact of cyber attacks, with the strongest uplift in reconnaissance and social engineering, and that it lowers the barrier to entry for less capable actors. The assessment has since been extended, judging AI-enabled malware in a ransomware attack on UK infrastructure highly likely by 2028. Check the current NCSC wording before quoting it to a board.
Does “highly likely by 2028″mean AI ransomware is not a problem yet?
No. “Highly likely”sits in roughly the 80 to 90 per cent band on the PHIA probability yardstick, so it is a planning assumption rather than a distant maybe. Separately, the AI-assisted parts of the attack chain, better lures and faster target triage, are happening now. The forecast concerns the payload, not the pressure.
Is AI-enabled ransomware a new type of malware or the same attack done faster?
Mostly the same attack, cheaper and faster, with a wider pool of people able to run it. Initial access remains phishing, exposed services and stolen credentials. Where something genuinely new appears, it tends to be inside the victim: agents and assistants with tool access that hand an intruder capability they used to have to earn.
Which AI systems in my organisation should be tested first?
Rank them on reachability, capability, data breadth, checkpoints, and reversibility and detectability. Anything that ingests untrusted text and can write, delete, pay or deploy goes to the front of the queue. A chatbot with no tools but access to the whole document store and no audit trail usually outranks a well-fenced agent.
Do standard ransomware readiness exercises already cover AI risk?
Partly. Backup immutability, segmentation and restore timing all still apply. What they usually miss is whether an AI service identity can reach backup storage, what a hijacked assistant can call, and whether injection through a ticket or a supplier document is detected at all. Add those questions to the exercise you already run.
