What AI and Cybersecurity Need Before We Can Trust Them
November 19, 2026 | Netlight, Stockholm
Insights by Stockholm MLOps
Five minutes before an important meeting, a salesperson received a verification link from what appeared to be a legitimate prospective customer. According to the anonymized case described by Gustav Thenander of Skeptiva, the conversation had developed like a normal enterprise sales process. The supposed buyer represented a defense organization, had asked reasonable questions, disappeared for roughly three weeks, and then returned ready to continue the discussion.
When the meeting was finally scheduled, the buyer requested identity verification shortly before it began. The explanation sounded responsible. Everyone should verify themselves before joining in order to protect against deepfake impersonation. The salesperson followed the process, but the security procedure itself was the attack.
The available event material does not establish the exact technical compromise or independently verify the incident, so it should not be treated as a forensic case study. What makes it important is the human and operational context. The employee was not reacting to an obviously suspicious message. He was operating inside a believable business relationship, under time pressure, with a potentially valuable customer, while being asked to follow what sounded like a legitimate security control.
Gustav summarized the underlying problem simply when he said, “We run on trust.” Modern business has to. As he also pointed out, the world would stop functioning if every employee had to independently verify every piece of information presented to them before taking any action.
That tension appeared throughout Stockholm MLOps #41. Alexandra Searle showed how security itself can become so complex that organizations struggle to understand ownership and priority. Hummam Wasfi questioned whether periodic risk reporting can keep pace with continuously changing technology and dependencies. Zack Korman challenged the cybersecurity industry's ability to prove that defensive products actually work. Mahan Tourkaman pushed the same problem into continuous red teaming and adaptive adversaries. Novo Nordisk and Netlight approached trust from a very different direction, showing how machine learning in healthcare has to survive external validation, clinical explanation, privacy requirements and deployment constraints before somebody should rely on it.
The talks covered very different systems, but the same question kept resurfacing. Before a person, company, clinician or automated system can safely act, what evidence is enough to justify that action?
That is the deeper cybersecurity problem exposed by #41.
Summary: Key Insights
The reported Skeptiva case showed why the most dangerous moment in a security system may occur when a legitimate business process and an attack become almost indistinguishable. Better employee awareness is useful, but it cannot be the entire defense when attackers can exploit timing, trust, commercial incentives and the language of security itself.
Alexandra Searle's perspective from Ericsson showed another side of the same problem. Security organizations can accumulate overlapping teams, strategies, taxonomies, controls, tools and reporting structures until the system becomes difficult to navigate. Adding AI agents on top of fragmented processes risks automating the fragmentation rather than resolving it.
Zack Korman raised an equally uncomfortable measurement problem. If an organization deploys a security product and no breach occurs, that does not automatically establish that the product caused the good outcome. Buyers need ways to distinguish genuine defensive value from claims that are difficult to falsify.
Mahan Tourkaman argued that adversarial testing will have to become more continuous as AI systems and attackers evolve. His discussion of swarms and adversarial stability was presented as an ambitious direction rather than a demonstrated autonomous security architecture, which is an important distinction to preserve.
The Novo Nordisk and Netlight presentation showed that trust in a machine learning system is not one thing. Clinicians need to understand individual predictions, payers need evidence that performance transfers to their own populations, builders need robust validation and data quality, and regulators need confidence in the intended use and deployment process.
Deployment choices also redistribute responsibility. A centrally operated healthcare application, a payer hosted Docker deployment and an EHR integration may use the same underlying modeling work, but they create very different tradeoffs around adoption, privacy, maintenance, governance and operational ownership.
The Mentimeter results added another important perspective from the room. Organizational readiness and culture was the largest reported barrier to moving AI and LLMs into production at 34 percent, while security, privacy and data sovereignty followed at 28 percent. Together, those two categories accounted for 62 percent of responses, reinforcing the idea that production AI is as much an organizational and trust problem as a technical one.
What This Event Was Really About
Stockholm MLOps #41 was framed around AI and cybersecurity, but the evening moved beyond the familiar argument that AI simply gives attackers and defenders more powerful tools. That dynamic was certainly present, yet the stronger story concerned assurance.
A salesperson needs evidence that a request from a prospective customer is genuine. A CISO needs evidence that a security product actually reduces risk. A risk owner needs information that still reflects the current environment rather than the environment six months ago. A red team needs tests that remain meaningful even when systems and attackers adapt to them. A clinician needs to understand why a particular patient received a prediction, while a payer needs to know whether the model continues to perform on a different population.
These are all trust questions, but they require different kinds of evidence.
That distinction is important because the word “trust” is often used too loosely in discussions about AI. A trustworthy model for a clinician is not automatically a trustworthy deployment for a hospital IT department. A compliant architecture is not automatically a clinically useful one. A successful penetration test does not guarantee security against every future attack. A security product that produces fewer alerts does not automatically make the organization safer.
The event repeatedly showed that evidence has to be matched to the decision being made.
The salesperson in Gustav's example did not need a general cybersecurity score. He needed useful information at one specific moment, shortly before a meeting, when challenging the buyer carried a commercial cost.
The doctor in the Novo Nordisk and Netlight case needed to understand why a patient received a particular risk estimate. The payer needed evidence that the ranking performance persisted in another population. The deployment team needed clarity about where data would be processed and who would operate the system.
The CISO considering a new security tool faces a different question again. Zack's argument was that the absence of an incident cannot by itself establish which defensive control caused the good outcome.
Seen together, #41 was less about whether AI is safe or unsafe and more about the mechanisms by which organizations decide what is safe enough to act on.
1. The Attack Looked Like Security
Gustav Thenander's reported case was memorable because the attacker did not begin by behaving like an attacker. The supposed buyer behaved like a buyer. There was an ordinary inquiry, a normal sales conversation, a long period of silence and then renewed commercial interest.
The attack only became actionable once the relationship had accumulated enough credibility. Shortly before the meeting, the salesperson received a verification request that was presented as protection against deepfakes.
Gustav described the timing directly: “Five minutes before the meeting, he gets sent a link.”
The technical mechanics of the compromise were not described in enough detail to identify the exact attack vector, and the article should not invent one. The human mechanism is much clearer. The salesperson had a reason to continue quickly and a reason not to create unnecessary friction with a potentially important customer.
That is why the usual instruction to verify everything has practical limits. Gustav argued that it is unrealistic to expect ordinary employees to withstand increasingly sophisticated attacks through vigilance alone. His more useful question was how organizations can give people the right information and the right tools at the moment a decision has to be made.
That changes how we think about the human role in security. Human behavior is not simply a weakness to be trained away. It is a design constraint that security systems have to account for.
A security control has to work when the employee is busy, when the request sounds reasonable, when the attacker has spent weeks building credibility and when delaying the process appears to have a business cost.
2. AI Changes the Economics of Social Engineering
Gustav also made an important point about scale. He said, “If you can automate sales outreach, you can automate an attack.”
AI does not need to invent an entirely new category of cybercrime in order to change the threat environment. It can alter the economics of attacks that already work.
A highly personalized social engineering campaign traditionally requires substantial human effort. The attacker has to research the target, maintain the conversation, adapt language, wait for the right moment and respond plausibly as circumstances change.
If AI can automate meaningful parts of that process, the cost of personalization falls and the number of potentially viable targets rises.
Gustav framed the question in economic terms when he asked, “What happens if the ROI shifts drastically for the attacker?”
That may be more useful than focusing only on spectacular new attack techniques. The underlying vulnerability may be familiar, while the cost structure changes enough to make persistent and personalized attacks practical at much larger scale.
The event did not provide a measured dataset showing how frequently AI generated social engineering is currently succeeding, and the large agent counts discussed in the talk were illustrative rather than observed attack volumes.
The useful production takeaway is therefore not that every phishing attack is now AI generated. It is that defenders need to consider what happens when the cost of sophisticated persuasion continues to fall.
3. Security Can Become Its Own Complexity Problem
Alexandra Searle approached cybersecurity from the inside of a large business rather than from the attacker's perspective.
Her Everyday Security material showed a landscape of overlapping disciplines, including IT security, product security, information security, physical security, privacy, business continuity, executive protection, insider threat prevention, resilience and crisis management.
Across those functions sat strategies, tactical plans, control frameworks, taxonomies, tools, personnel, reporting processes and risk identification mechanisms. The issue was not that any of those functions are unnecessary. The problem is what happens when each function evolves its own vocabulary, processes, reports and ownership structures.
An organization can therefore have a large number of highly competent security specialists and still create an operating environment that is difficult for the business to navigate.
This becomes particularly relevant as organizations introduce AI agents into security operations. Automation applied to a clear process can remove repetitive work. Automation applied to a fragmented process can simply execute the fragmentation faster.
If ownership is unclear, requirements conflict, evidence exists across multiple systems and critical knowledge sits inside individual people, an AI agent does not automatically solve the architecture. It becomes another actor operating inside it.
This creates an important counterweight to the assumption that the cybersecurity answer to AI is necessarily more AI. In some situations the first engineering task may be organizational simplification. Ownership has to become clearer, duplicated reporting has to be reduced and the organization has to agree which risks deserve priority before deciding what should be automated.
The event did not present a quantified before and after study showing that organizational simplification directly reduced cyber incidents. The evidence was structural rather than experimental.
Even so, Alexandra's argument raises an important production question. If AI increases the speed at which work moves through the organization, what happens when the underlying operating model is already too complex?
4. How Do You Know a Security Product Works?
Zack Korman focused on one of cybersecurity's most uncomfortable measurement problems: how to prove that the defense actually caused the protection it claims.
His memorable analogy involved a box with a light that supposedly activates whenever a unicorn enters the room. If the light never turns on and no unicorn appears, the absence of a unicorn does not tell us whether the box worked perfectly or whether there was never a unicorn to detect.
The analogy is absurd, but the measurement problem is real.
A company can deploy a security product and experience no breach. That does not isolate the causal contribution of the product. Perhaps the organization was not attacked, another control stopped the attack, the attack went undetected or the product genuinely worked.
The outcome alone cannot tell us which explanation is correct.
Zack extended that critique toward simulations, benchmarks, vendor claims and procurement. Security teams need tests because they cannot buy products entirely on trust. Yet a product can potentially become very good at performing against the tests used to evaluate it without becoming equally effective against an adaptive attacker outside the test.
That does not make benchmarks useless. It means the benchmark itself has to be treated as an imperfect model of the thing we actually care about.
The event did not provide a universal solution to cybersecurity efficacy measurement, but it raised a valuable standard for evidence. If a vendor says a control works, the buyer should be able to ask what exactly was measured, against which threat, under which conditions and how the result relates to a real reduction in risk.
5. When the Red Team Becomes Part of the System
Mahan Tourkaman extended the evaluation problem into more speculative territory.
His argument was that adversarial testing will have to become increasingly continuous as AI systems and attackers evolve. Rather than relying only on occasional assessments, future defensive systems may use automated agents to keep probing for weaknesses and continuously test whether assumptions still hold.
Mahan summarized his view by saying, “Red teaming is here to stay.”
He also described a longer term vision in which this kind of adversarial testing becomes something like an immune layer for digital systems. That is a provocative research direction, but the evidence from #41 does not establish that such an autonomous defensive architecture has been built or validated. No working swarm, coverage benchmark or production control architecture was demonstrated.
Mahan was open about the theoretical nature of the challenge. He asked, “What long-term guarantees?” and suggested that stronger mathematical theories may eventually be required.
That uncertainty is valuable because it separates an interesting research problem from an established security practice.
The connection to Zack's talk is particularly interesting. If defenders optimize against a fixed benchmark, the benchmark can become another target. A sophisticated attacker is not required to behave like the benchmark used in procurement or testing.
Continuous adversarial evaluation is therefore attractive because it can change the conditions under which the defense is tested. However, the more autonomous the offensive testing becomes, the more carefully the testing system itself has to be controlled.
Security automation eventually creates another security system that must be secured.
6. Risk Information Has an Expiration Date
Hummam Wasfi approached the evening from the perspective of organizational risk and resilience.
A modern risk environment changes continuously. Technology changes, suppliers change, dependencies change, vulnerabilities emerge, AI systems are introduced and data flows evolve.
Many organizational risk processes still operate on much slower cycles through spreadsheets, SharePoint documents and periodic review processes. The problem is not that those mechanisms are inherently useless. It is that the representation of risk can become stale before the next formal review occurs.
Hummam contrasted that periodic approach with the possibility of more continuous risk intelligence, where technical, supplier and organizational signals can be combined into a more current view of exposure.
That idea sounds attractive until it is placed next to Alexandra's argument about organizational complexity.
More information does not automatically produce more clarity. A continuously refreshed system that generates endless signals, scores and alerts can reproduce the same complexity problem at greater speed.
The useful goal is therefore not simply real time reporting. It is current risk information connected to clear responsibility and action.
The important questions become operational. Who owns the response? Which evidence is enough to justify escalation? Which risks deserve attention first? What can an automated system act on, and what still requires explicit human approval?
A continuously refreshed dashboard without those answers can become another layer of noise.
7. Healthcare Shows What Stronger Evidence Can Look Like
The joint Novo Nordisk and Netlight presentation brought a very different domain into #41, but it provided some of the evening's clearest examples of what verification looks like when a model may eventually influence consequential decisions.
The team was working with obesity related comorbidities and the question of which patients may face greater future disease risk.
Instead of building one generic classifier, Lorena Torres Lahoz explained that the team used a different model for each obesity related comorbidity. They also used survival analysis because the relevant question was not simply whether a patient would eventually receive a diagnosis, but how the risk evolved over time.
Healthcare datasets also contain censoring. A patient who has not experienced an event by the end of the observation period cannot automatically be treated as someone who will never experience it.
The modeling choice therefore followed the clinical question and the structure of the data rather than forcing the problem into a simpler binary classification framework.
The project then moved beyond model development and asked what evidence different stakeholders would need before relying on the predictions.
That is where the healthcare example becomes particularly relevant to the cybersecurity discussion.
8. One Trust Score Is Not Enough
Anna Lewicka framed one of the healthcare questions from the clinician's perspective when she asked why a particular patient had been given a particular number.
For that question, the team used SHAP values to help explain which features contributed to a prediction.
However, explanation answers only one kind of trust question. A model can produce a clear explanation for a prediction and still perform poorly on a different population.
That is why payers required another type of evidence.
The team discussed validation on an independent dataset from a demographically different population. They reported approximately 4.5 times enrichment over random selection internally and almost four times enrichment in the external dataset.
The team described the external result by saying, “We were still doing pretty well, almost four times better than random selection.”
This is useful reported evidence that the ability to rank higher risk patients remained relatively strong when the population changed. It should not be interpreted as meaning treatment was four times more effective, four times as many conditions were prevented or reimbursement decisions based on the model improved outcomes.
The distinction matters because different stakeholders mean different things when they say they need to trust the model.
A clinician may need an explanation.
A payer may need evidence of transportability.
The model team needs data quality and validation.
A regulator needs confidence in the intended use and the way the system handles risk.
Trust therefore cannot be represented by one universal score. It has to be connected to a particular decision and a particular stakeholder.
9. Deployment Changes Who Owns the Risk
The healthcare presentation became even more relevant to production engineering when the team moved from modeling into deployment.
Three broad delivery paths were considered.
A centrally operated standalone web application gives the solution provider substantial control over the environment. It also creates another login and another workflow for clinicians, which can become an adoption problem even if the software itself is technically straightforward.
A payer hosted Docker deployment moves much more control to the local organization. As the team explained, the payer gains greater autonomy, but the operational burden also moves with the deployment. Somebody has to operate, maintain, secure and support the environment.
The third option, EHR integration, brings the system closer to the clinician's existing workflow and may therefore improve adoption. At the same time it increases dependencies on external systems, integration partners and governance processes.
None of these architectures makes compliance or security disappear.
They redistribute responsibility.
Where the system runs affects who handles the data, who operates the infrastructure, who investigates failures, who maintains the application and who is accountable when something goes wrong.
This is why deployment architecture in regulated AI is not merely an infrastructure choice. It is also an operating model choice.
10. Compliance Works Better When It Arrives Before Deployment
One of the more useful comments from the Novo Nordisk and Netlight team concerned when compliance enters the development process.
Their approach was to think about it from the beginning and then carry those requirements through the development process rather than treating compliance as a final approval step.
That matters because regulatory and privacy requirements can change the technical architecture long before deployment.
Whether healthcare data is anonymous or pseudonymized affects how it can be processed. The intended purpose of the system influences the regulatory path. A payer hosted deployment creates different responsibilities from a centrally managed application. EHR integration introduces new dependencies and controls.
These are architecture questions as much as compliance questions.
The same principle applies in cybersecurity.
Governance is most useful when it shapes the architecture while the architecture is still being designed. If a system has already been built around assumptions that data may move freely, agents may act without explicit mandates or every component may access every tool, security becomes much more expensive to retrofit.
The strongest form of governance is therefore not a checklist added at the end. It is a set of constraints reflected in how the system is built.
11. The Human Is Not the Weakest Link
A familiar cybersecurity phrase describes people as the weakest link.
The evidence from #41 suggests that this is too simplistic to be useful.
The salesperson in Gustav's example made a decision inside a deliberately manipulated business context. Clinicians in the healthcare example provide domain judgment that cannot simply be replaced by increasing model accuracy. Security owners still have to decide which actions autonomous systems are allowed to take. Risk leaders have to decide which exposures deserve priority.
Alexandra's organizational complexity also showed that humans are often navigating systems whose structure makes correct behavior unnecessarily difficult.
This is not an argument against automation. It is an argument for being precise about where human judgment belongs.
Some human involvement exists because the system is poorly designed. Asking employees to manually verify every communication is not a scalable security architecture.
Other forms of human involvement are intentional controls. A physician deciding whether a prediction makes clinical sense, or a security owner authorizing a high impact action, may be exactly where human judgment provides the most value.
The important question is therefore not whether a human remains somewhere in the process. The more useful question is what decision the person is expected to make, what evidence is available at that moment and whether they actually have the authority, context and time required to make the decision well.
Insights by Stockholm MLOps: Signals From the Room
The Mentimeter results add an important layer because they show what practitioners in this particular room reported as their biggest barriers to moving AI and LLMs into production.
Organizational Readiness and Culture was the largest category at 34 percent, while Security, Privacy and Data Sovereignty followed closely at 28 percent. Data Quality and Pipeline Hygiene, Model Evaluation and Observability, and High Compute Costs or ROI Justification each received 10 percent, while Governance and Regulatory Compliance received 9 percent.
These results should not be presented as an industry survey. They represent this Stockholm practitioner audience at this event.
The pattern is still striking.
The two biggest categories were not model performance, GPU availability or inference speed. They were organizational readiness and security, privacy and sovereignty. Together, they represented 62 percent of responses.
That creates a strong connection between the room and the talks on stage.
Alexandra described organizational complexity. Gustav showed how a reasonable employee can be placed inside a decision context where generic security advice becomes insufficient. Hummam focused on making fragmented risk information current enough to support decisions. Novo Nordisk and Netlight showed how privacy, validation, regulation and workflow constraints shape architecture before deployment.
The room was pointing toward the same broader problem from another direction. Getting AI into production is not only about whether the technology is capable enough. Organizations also need the confidence, responsibility structures and operating model required to use it.
Gustav's informal show of hands around email use reinforced that point in a different way. There was no count, so it should not be turned into a statistic. The value of the exercise was to remind the room that modern organizations function because people make trust decisions constantly.
Security cannot remove trust from business. It has to make that trust more defensible.
Only Said Here: What Practitioners Actually Taught Us
Some of the most useful observations from #41 came from small details that would rarely appear in a generic article about AI and cybersecurity.
The reported attack did not begin with urgency. The fake buyer spent weeks behaving plausibly before the decisive request appeared. The malicious step was then presented as protection against deepfakes, which meant that security language itself became part of the social engineering pretext.
The timing also mattered. The link arrived shortly before an important meeting, when slowing the process down conflicted directly with the commercial objective.
Alexandra's Everyday Security view made another normally invisible problem visible. Similar categories of planning, reporting, control and tooling were repeated across numerous security disciplines, showing why adding another layer of automation does not automatically reduce complexity.
Zack's analogy exposed a measurement problem that many buyers would rather avoid. An absence of bad outcomes does not prove that a particular control caused those outcomes to disappear.
Mahan raised the related problem on the offensive side. If a security team becomes too successful at optimizing against a known benchmark, the benchmark can become another target rather than a reliable measure of resilience.
The healthcare team provided some of the evening's strongest positive evidence. They used separate survival models for different clinical outcomes and tested whether risk ranking performance persisted on an independent population.
They also showed that a payer hosted Docker image, which sounds at first like a technical packaging decision, can actually become a decision about sovereignty, maintenance, compliance and organizational responsibility.
Those details are what make the event useful. They move the conversation away from the generic statement that AI will change cybersecurity and toward the specific ways practitioners may have to design systems differently.
The Verification Problem
The talks at Stockholm MLOps #41 covered social engineering, security organization design, risk intelligence, defensive product efficacy, autonomous red teaming, European compute and regulated healthcare ML.
The technologies were different, but the same deeper question appeared repeatedly: how do we know enough to act?
The salesperson has to decide whether a verification request is genuine. The security buyer has to decide whether a defensive product really reduces risk. The red team has to decide whether a benchmark reflects resilience or merely performance against a known test. The risk leader has to decide whether an assessment still describes the organization as it exists today. The clinician needs to understand why a patient received a prediction, while the payer needs evidence that the model remains useful on its own population.
Those decisions cannot all use the same kind of evidence.
That may be the most important conclusion from #41.
Trustworthiness is not one metric. A SHAP explanation cannot replace external validation. A penetration test cannot guarantee protection against every future attacker. A compliance checklist cannot prove that a system is resilient. A security warning is not useful if it arrives after the decision has already been made. An AI agent cannot resolve an organizational structure that has never established what the agent is authorized to do.
The objective is therefore not to eliminate uncertainty, because that is unlikely to be possible.
The objective is to design systems where the evidence, responsibility and controls are appropriate to the decision that actually has to be made.
That is what moves security from something added after development into something that belongs inside the production architecture.
What Stockholm MLOps #41 Signals
AI is increasing capability on both sides of cybersecurity. Attackers can potentially automate more research, persuasion and experimentation, while defenders can automate more monitoring, testing, analysis and response.
The evidence from #41 suggests that capability alone will not determine which side benefits most.
The harder problem is assurance.
Organizations need to know when their controls are working, when risk information has become stale, whether tests survive adaptive adversaries, how employees can act safely without becoming security analysts, and whether machine learning systems remain reliable when they leave the environment in which they were developed.
They also need to understand the operational consequences of control. Local deployment can increase autonomy and data control while transferring more maintenance and compliance responsibility to the organization. More automation can increase speed while also increasing the number of systems that require permissioning, evaluation and oversight.
The Mentimeter results make this difficult to ignore. Organizational readiness was the largest reported production bottleneck in the room, while security, privacy and sovereignty came second.
The message from #41 is therefore not that cybersecurity should slow AI down. It is that production AI requires trust to become something organizations can deliberately design, test and assign responsibility for.
The broad question, “Can we trust AI?” is too vague to be useful.
A better question is:
What evidence would justify trusting this system for this decision under these conditions?
That question is harder, but it gives engineers, security teams, clinicians and business leaders something concrete to work on.
Speakers and Organizations
Alexandra Searle, Ericsson presented a business perspective on cybersecurity complexity and the challenge of prioritizing risks across overlapping controls, processes and security disciplines as AI increases the pace of change.
Hummam Wasfi, Nordic Institute for AI Risk & Resilience explored emerging AI risk, governance, resilience and the need for risk information that can evolve with changing technologies, suppliers and dependencies.
Zack Korman, Embroidery challenged cybersecurity's ability to demonstrate that security products and controls are actually effective and questioned how buyers distinguish meaningful evidence from difficult to falsify claims.
Gustav Thenander, Skeptiva focused on sophisticated social engineering, the changing economics of AI assisted attacks and the limits of relying on perfect human vigilance.
Mahan Tourkaman, Speedrun explored continuous red teaming, offensive AI systems and the longer term research question of whether stronger forms of adversarial stability can be created.
Anna Lewicka and Lorena Torres Lahoz, Novo Nordisk, together with Alizé Papp of Netlight, presented Future Health Models, showing how predictive ML in regulated healthcare has to combine appropriate modeling, external validation, explanation, confidentiality, compliance and deployment architecture.
Petra Dalunde, Sweden AI Factory presented pathways into Swedish AI compute and support for researchers, startups, companies and public organizations.
Stockholm MLOps #41 took place at Netlight in Stockholm to a full house despite snow in Stockholm and Slush taking place in Finland.
Related Insights
Stockholm MLOps #40: The Production Decision Starts With Constraints
#40 showed that production AI decisions increasingly depend on measurable quality, policy, verification and workload constraints. #41 continues that discussion from a security and assurance perspective, where the question becomes what evidence is sufficient before a person or system should act.
Stockholm MLOps #30: AI in Production, Healthcare
The healthcare work at #41 continues a theme that has appeared repeatedly in Stockholm MLOps healthcare discussions. The difficult part is rarely only the predictive model. Measurement, workflow integration, trust, deployment and responsibility determine whether the model can create value in practice.
Sovereign AI & Control AI
#41 adds another dimension to control. Data sovereignty and local deployment can reduce some forms of exposure, but the healthcare and security discussions showed that greater control can also transfer more operational responsibility to the organization running the system.
Event Details
Stockholm MLOps #41
AI & Cybersecurity
November 19, 2026
Netlight, Stockholm
Featuring Ericsson, Nordic Institute for AI Risk & Resilience, Embroidery, Skeptiva, Speedrun, Novo Nordisk, Netlight and Sweden AI Factory.