A financial controller exports a sales report containing customer data on a Friday evening and sends it to a personal address to finish the analysis over the weekend. Two weeks later, the account of another person in the same department signs in at night from an unusual location and downloads an archive of contracts. The first situation is a well-intentioned mistake, the second is probably a stolen password. In both cases the data left the company through the hands of someone the organisation had given access to, and in both cases the security systems saw an authorised user.
This article is for CISOs and risk owners who have to answer the board and the auditor: what are we doing about risk that comes from inside? The definition of insider threat and its types are covered in our glossary entry on insider threat. Here we focus on practice: how to recognise the scenario, which signals to collect, how to split roles between security, HR and legal, how to control access and data, and how to check whether the programme actually works.
How does insider risk differ from an external threat?
Insider risk starts with access that already exists, not with a break-in. An external attacker has to get past the defences; an insider — an employee, a contractor, a supplier’s staff member — is already on the right side of them and uses permissions that were granted deliberately. That is why the classic question “has an outsider got into the network?” is not enough. You also have to ask whether a person with access is doing something with data that their role does not require.
The US agency CISA, in its Insider Threat Mitigation Guide (September 2026 edition), defines an insider broadly: anyone who currently holds, or previously held, authorised access to the organisation’s resources or knowledge about them. The word “previously” matters — a former employee with an account that was never disabled, or a contractor after the end of an assignment, still fits the definition. The threat is the possibility that such a person uses that access or knowledge to harm the organisation, whether on purpose, through negligence or by accident.
The difference has three practical consequences for the security team:
- The signal is weaker. The insider’s sign-in is valid, the permissions comply with policy, the network traffic looks normal. The deviation shows only in what someone does with data, and when.
- The context sits outside IT. A resignation, a change of role, a dispute over a bonus or the end of a contract are HR information. Without it, the same file export is routine one day and a warning sign the next.
- The response concerns a person. The answer to a break-in is to cut the attacker off. The answer to an insider signal may be a conversation, a reminder of the rules or disciplinary proceedings — and that needs HR and the legal team.
That is why the discipline is called insider risk management rather than just threat detection. Management covers deciding which data to protect most strongly, who reaches for it, which signals count as significant and who takes the decision when they appear.
How big is the problem of data loss involving people with access?
Big enough that the financial impact of the most serious incidents is counted in millions. According to the 2025 Data Security Report, produced by Fortinet and Cybersecurity Insiders from a survey of 883 IT and cybersecurity professionals, 45% of organisations saw lost revenue or a financial loss as a result of their most serious data exposure incident, and 41% put their losses at between USD 1 million and USD 10 million.
These figures need to be read with three caveats that the report itself gives. First, the answers are respondents’ own declarations, not audit results. Second, the sample is made up overwhelmingly of organisations with more than 250 employees — smaller companies are a negligible share of it, so the results cannot be transferred to businesses with a few dozen staff. Third, the report describes incidents involving insiders in a broad sense: mistakes and compromised accounts included. Figures on how often such incidents occur and on the share of carelessness are collected in our insider threat glossary entry — we do not repeat them here.
More important than the scale is the conclusion for the budget. If the most serious incident can cost as much as a multi-year data protection programme, the conversation with the board is not about whether to invest but about where to start. The answer is rarely “with a tool”. It is usually: with the data whose loss would hurt most, and with the people who have access to it.
In short: data loss involving insiders is a financial problem, not just a compliance one. Industry report figures show the scale in large organisations — in a mid-sized company the numbers will differ, but the mechanism is the same.
How do you tell a mistake, abuse and a compromised account apart before a decision is made?
By the effects and the context, not by intent — intent is not visible at the moment of the first signal. CISA divides insider threats into unintentional ones (negligence and accident) and intentional ones, and separately points to collusion with outsiders and to threats from third parties, meaning contractors and suppliers. The same guide notes that most recent incidents are unintentional, although a single intentional incident usually costs more.
In practice the security team sees the event first, not the motive. It is therefore better to plan the response around a scenario that can be recognised from its traces than around a “malicious” or “careless” label. The table below is a working map we use when designing monitoring rules:
| Scenario | First visible trace | Control that stops it | Who decides on next steps |
|---|---|---|---|
| Mistake when sharing | a file with customer data sent to the wrong address or shared with “anyone with the link” | email and cloud DLP with a warning at the moment of sending, safe default sharing settings | the line manager and the data owner; security only for repeat cases |
| Bypassing rules for convenience | company data copied to a personal cloud drive, personal email or a public AI tool | blocking or notification for personal accounts, an approved alternative (a company tool) | security together with the line manager; HR for persistent rule-breaking |
| Compromised account | sign-in from a new device or location, unusual hours, bulk downloads | multi-factor authentication, sign-in analytics, fast session suspension | security on its own — this is an incident, not an HR matter |
| Deliberate exfiltration before leaving | a rise in downloads and archive compression after a resignation | correlating events with HR information, restricting access during the notice period | a cross-functional team including HR and a lawyer |
| Abuse of administrative privileges | permission changes, disabled logging, access to data outside the scope of duties | PAM with session recording, separation of duties | a cross-functional team; often the board |
The most important column is the last one. A mistake and a compromised account do not call for a disciplinary conversation, and trying to handle them as an investigation destroys trust in the programme. Deliberate scenarios, on the other hand, cannot be left to a single administrator, because they require evidence to be preserved and a decision that IT should not take alone.
The boundaries between scenarios are fluid, which is why we treat the classification as a working hypothesis. The team confirms or changes it once the context is gathered — just as a SOC analyst changes the priority of an alert after checking what lies behind it.
Which signals in data and behaviour point to a data leak?
What tells you most is not individual events but their sequence: collecting data, preparing it for removal and the removal itself, within a short time, by one person. The signals are worth dividing into three groups, because each is gathered by a different team.
Data signals come from systems that see files and their movement: DLP, cloud drive, email and application logs. Typical examples are downloading a large number of documents from a repository the person rarely uses, packing files into password-protected archives, renaming files to something less obvious, copying to removable media or uploading to a personal account in a cloud service. Each of these actions can be justified. Together, and close together in time, they form a pattern worth checking.
Identity and access signals come from authentication and permission management systems. A sign-in from a new country, from an unknown device or at an hour when the person never works more often points to a compromised account than to a dishonest employee. This group also includes requests for permissions the role does not need and attempts to reach resources someone has no right to.
Context signals sit with HR and line managers: a resignation, the end of a contract, a move to another department, a pay dispute, a missed promotion. CISA calls these professional stressors and stresses two things. First, behaviour matters, not motivation. Second, the presence of even many indicators does not prove that someone is a threat — and their absence does not prove that they are not.
This leads to a principle we propose when configuring rules: a context signal does not trigger monitoring of a person; it only changes the weight of technical signals. A resignation on its own is no reason to analyse an employee’s activity. But downloading a project archive in the week after a resignation should reach an analyst faster than the same event for someone without that context.
Signals only make sense against a norm. The CERT Common Sense Guide to Mitigating Insider Threats (seventh edition, Software Engineering Institute, 2022), in one of its 22 best practices, recommends establishing a baseline of normal network and device behaviour. Without a baseline every large export looks suspicious, and the team soon stops reacting to alerts. User and entity behaviour analytics, or UEBA, automates exactly this comparison with the norm.
How do you build an insider risk management programme with HR, legal and IT?
As a permanent, cross-functional process with a single owner, not as a security department project. NIST SP 800-53 Rev. 5, in control PM-12, requires a programme with a cross-discipline team for handling insider incidents and points out that its strength lies in combining technical and non-technical information — including HR records. CISA adds a list of elements that recur in mature programmes: a stakeholder committee, a prioritised list of critical assets, a central place for collecting and analysing information, a threat management team backed by leadership, an incident response plan and training.
In an organisation of a few hundred people, this model comes down to a handful of roles:
| Role | What it contributes | What it does not do |
|---|---|---|
| Programme owner (CISO or chief risk officer) | sets scope, priorities and metrics; reports to the board | does not run investigations alone |
| Security / SOC | collects and correlates technical signals, analyses events | does not assess the employee’s personal situation |
| HR | passes on HR events (departures, transfers, contract ends), takes part in conversations | does not receive raw activity logs |
| Legal / data protection officer | sets the limits of monitoring, the legal basis and the rules for preserving evidence | does not choose the tools |
| Data and system owners | identify critical data, approve exceptions to policies | do not analyse alerts |
The split in the “what it does not do” column is as important as the tasks. It protects employees from a situation in which one person sees both their logs and their personal problems, and it protects the organisation from decisions taken without the right competence.
The team also needs a single place where a case lives from the first signal to closure. It does not have to be a separate system — a case register with restricted access, recording signals, findings, decisions and the people who took them, is enough. Such a register serves two purposes at once. It lets the team notice that three minor events from different months concern the same person, and it gives the organisation evidence that every decision regarding an employee had a documented basis. Only the team has access to the register, and cases closed without a confirmed threat are marked as such so that they do not weigh on anyone in the future.
Privacy is a condition for the programme to work, not an obstacle. In Poland, monitoring of business email is governed by Article 22³ of the Labour Code (consolidated text: Journal of Laws 2026, item 1245). The provision allows it to be introduced where this is necessary to organise work so that working time is used in full and the tools provided to the employee are used properly, and it stipulates that monitoring must not breach the confidentiality of correspondence or other personal rights of the employee. The same rules apply accordingly to other forms of monitoring. The purposes, scope and method of monitoring are set out in a collective agreement, the work regulations or a notice, and employees are informed of it no later than 2 weeks before it starts.
For an insider risk programme this leads to three practical decisions. First, the scope of the data collected is described in the work regulations before the tool is deployed, not afterwards. Second, the analysis concerns events involving company data, not the content of private correspondence. Third, access to the details of a specific person is opened only by a team decision, not by an administrator’s curiosity. Whether a given tool and its settings fall within the purposes of the provision is for the legal team to assess — it is worth planning this step before the purchase.
The programme also needs a reporting culture. CISA points out that employees are the first line of spotting a problem, provided they know whom to report a concern to and trust that the report will be treated confidentially. A simple channel for reporting mistakes — “I sent the file to the wrong recipient” — shortens response time more than the best detection rule.
How do you control access throughout employment, not just when someone leaves?
By reviewing permissions at every change of role, not only at joining and leaving. The most neglected moment is a transfer: an employee changes department, receives new permissions, and the old ones stay. After a few years and three changes of position, they have access to data from four departments. This is permission drift, and it means a single compromised account opens far more than it should.
The foundation is the principle of least privilege — NIST SP 800-53 describes it in control AC-6 as allowing only the access needed to carry out assigned tasks. Control PS-5 in the same catalogue requires that, on a transfer or change of position, the organisation confirm whether existing access is still needed and adjust it to the new role. CERT, in its guide, links least privilege with separation of duties: an operation with serious consequences should not depend on one person.
It is worth breaking the employment cycle down into four control points:
- Joining. Permissions are granted from the role, not “the same as the predecessor”. The account has an owner on the business side.
- Change of role. HR automatically notifies the transfer, and the owner of the new role confirms the scope. Old permissions expire if nobody confirms them.
- Notice period. Access to critical data is limited to what the handover requires, and signals from this period carry more weight.
- Leaving. Control PS-4 in NIST SP 800-53 requires access to be disabled within a time period defined by the organisation, credentials to be revoked, an exit interview that reminds the person of their obligations, equipment to be retrieved and the organisation to retain access to information the departing person managed.
The last point of PS-4 is often skipped. The mailbox, drive and documents of a departing employee must not disappear together with the account — they are often the only material that makes it possible to establish what happened to the data in the final weeks.
People outside the payroll need separate attention. Contractors and suppliers’ staff often have access comparable to employees but do not go through HR processes, so no HR event will close their accounts. An external account should have an expiry date and a guardian who is responsible for closing it. How to translate this into concrete decisions about the account and the session is described in our article on third-party vendor access. Administrator accounts — your own and other parties’ — are best handled by privileged access management (PAM), because it provides session recording and separates passwords from people.
How do you protect data in email, the cloud and AI tools?
By following data across every channel it can leave through, rather than securing one channel very well. An employee who is blocked from sending a file by email usually has several other routes within reach: a personal cloud drive, a messaging app, the browser, removable media, a public generative AI tool. If the policy only protects email, the mistake will move to a channel nobody is watching.
This is exactly what the tool landscape looks like today in many large organisations. According to the 2025 Data Security Report (Fortinet and Cybersecurity Insiders), endpoint DLP is in use at 48% of respondents, DLP for email at 46%, cloud DLP or CASB at 41%, network DLP at 37%, and tools for data classification or DSPM at just 28%. Each layer protects its own fragment, but none is universal. The last figure is the most telling: an organisation that does not know where its sensitive data is stored cannot effectively watch how it moves.
That is why data protection in an insider risk programme starts with classification, not with blocking:
- Critical data first. The first of the CERT best practices is to know and protect critical assets. In practice it is a short list: customer data, contracts, designs and source code, financial data before publication, HR data. The strictest rules are set for these.
- A label on the document. Classification makes sense if it travels with the file — a sensitivity label applied in the office suite is read by email and cloud DLP.
- Telling company and personal accounts apart. Uploading a file to the company cloud drive is work; uploading the same file to a personal account in the same service is exfiltration. The rules have to see this difference.
- A warning instead of a block wherever possible. A message at the moment of action (“this file contains customer data, you are sending it outside the company”) stops most mistakes without stopping work. Blocking is reserved for critical data and for channels with no business justification.
Generative AI tools have become a separate channel. Pasting part of a contract into a public chat has the same effect as sending it to a stranger, although employees usually do not see it that way. It is not worth starting with a ban that will push the problem into shadow AI. A more effective path is a company AI tool with clear rules, detection of public tool usage and a warning when someone tries to paste classified data.
Email remains the channel where mistakes happen most often for purely mechanical reasons: address autocomplete, reply-all, an attachment with the wrong version of a file. Mechanisms that check content and attachments before sending, enforce encryption of messages containing sensitive data and make it possible to reconstruct what went out and to whom work best here. We write more about Data Loss Prevention technology itself and about the strategy for deploying it in our article on data leak protection.
How do you monitor activity and respond without turning the company into a surveillance system?
By monitoring events involving company data according to transparent rules, and raising the level of detail only when a specific signal justifies it. A programme that records everything about everyone from day one is expensive, hard to defend legally and — paradoxically — less effective, because analysts drown in data.
The expectations of security teams themselves point in a similar direction. According to the 2025 Data Security Report (Fortinet and Cybersecurity Insiders), respondents expect three things above all from next-generation tools: behavioural analytics working in real time (66%), visibility into data from the first day of deployment (61%) and the ability to keep unsanctioned AI and SaaS applications in check (52%). It is a shift from static rules such as “does the file contain a national ID number” to the question “is this person behaving differently than usual”.
The response is worth designing as a ladder with transparent thresholds, not as a single “block” button:
- Observation. A single low-weight event goes into the log and counts towards the baseline. Nobody analyses the specific person.
- Interaction with the user. For a medium-weight event, the system or an analyst sends a message: a reminder of the rules, a request for justification, a pointer to an approved alternative. Most mistakes end at this stage.
- Team analysis. A sequence of events, or a technical signal combined with a context signal, goes to an analyst who gathers the facts and assesses the scenario.
- Cross-functional decision. When the analysis points to a deliberate scenario, the decision on next steps — a conversation, restricting access, preserving evidence — is taken by a team that includes HR and a lawyer.
- Immediate response. A compromised account or ongoing exfiltration of critical data justifies immediate suspension of the session or account by the security team, without waiting for a meeting.
The ladder has two advantages. An employee who made a mistake deals with the system or their manager, not with the security department. The analyst works on the few cases that really need attention instead of hundreds of alerts about sent attachments.
Monitoring only makes sense if someone responds to signals quickly and at any hour — compromised accounts do not respect working hours. Signals from DLP, identity systems and behavioural analytics are best sent to the team that already correlates security events, namely the SOC. There, an insider signal can be set against other events — for example an earlier phishing attempt on the same account — which often settles whether you are dealing with an employee or with an attacker wearing their identity.
Every case should end with a record: what was detected, how it was classified, which decision was taken and why. CISA recommends reviewing the programme after each case is closed, and CERT devotes the last of its 22 practices to learning from past incidents. Without a record, every new case starts from scratch.
How do you measure whether an insider risk programme works?
By the speed and accuracy of the response and by closing gaps in access — not by the number of alerts or blocked files. The number of alerts rises after every new rule and falls after every suppression, so it says more about the configuration than about the risk. The board and the auditor need measures that show whether the organisation handles data and access faster and more accurately.
The set we suggest starting with:
| Metric | How to calculate it | What to watch out for |
|---|---|---|
| Time to remove access after leaving | from the end of the engagement in the HR system to the deactivation of the last account, separately for privileged accounts | an average hides single accounts forgotten for months — look at the worst case |
| Accounts without an owner | the number of active accounts (including external and technical ones) for which no responsible person can be named | technical accounts need a separate list, otherwise they always come out “without an owner” |
| Permissions confirmed at a change of role | the share of transfers after which the owner of the new role confirmed the scope of access | bulk confirmation without review gives a high score and zero value |
| Time from signal to decision | from the event that crossed the analysis threshold to the classification of the scenario | measure separately for compromised accounts and HR cases — they move at different speeds |
| Accuracy of analysis | the share of cases passed to the team that turned out to be justified | very high accuracy may mean the thresholds are too high and minor cases slip through |
| Mistakes stopped at the moment of action | the number of warnings after which the user gave up sending or chose the approved alternative | rises after a new rule is deployed; look at the trend, not the absolute value |
| Channel coverage | which exit channels (email, cloud, removable media, browser, AI) have a rule for critical data | ”we have DLP” without this map does not answer the auditor’s question |
Each of these measures has an owner. Time to remove access belongs jointly to HR and IT, accuracy of analysis to the security team, channel coverage to the programme owner. A quarterly review is enough to see the trend; a review after every serious case lets you check whether the metrics describe reality at all.
The report for the board should not contain names or case descriptions. Three elements are enough: the trend of the metrics from the table, a list of gaps in channel and access coverage with a plan to close them, and a short note on the cases closed in the period — how many there were, which scenarios they concerned, whether they required a change of rules. A report built this way answers the question the board really asks: does the organisation have control over who does what with its data.
It is also worth testing the programme in exercises. CISA recommends a progressive approach: from discussion-based exercises in which the team walks through a scenario around the table, to full operational exercises. A good first scenario is an employee who resigns on Friday and downloads a project archive over the weekend — walking the path from signal to conversation reveals gaps faster than any document audit.
In short: measure time and accuracy, not the number of alerts. The metric that improves fastest in a new programme is usually time to remove access — and it is the one auditors ask about most often.
What does reducing insider risk look like with FortiDLP and FortiMail as an example?
FortiDLP and FortiMail cover two of the layers described above: user behaviour on the device and in cloud applications, and the email channel. This is an implementation example, not the only way — the mechanisms described can also be built from other tools, provided they perform the same functions.
FortiDLP — data and behaviour on the device and in the cloud. According to the vendor’s documentation, FortiDLP brings together in one solution data loss prevention, insider risk management, data protection in GenAI and SaaS tools, and user education. A lightweight agent protects the flow of data regardless of whether the employee is in the office or working remotely, and the solution supports the Microsoft 365, Google Workspace and Box cloud drives. Machine learning built into the agent establishes a baseline of the individual user’s behaviour, and the recorded activities are mapped to the MITRE knowledge base of insider threat tactics and techniques.
In the context of this article, three functions matter most. The Data Origin and Data Lineage mechanisms make it possible to enforce policies according to where the data came from — for example the HR system or a source code repository — and according to whether the data leaves through a company or a personal account. The Sequence Detection function correlates actions from the stages of collecting data, covering tracks and exfiltration, which is exactly the sequence described in the section on signals. The response is graduated: from notifying the user via Microsoft Teams or Slack, through capturing file and clipboard artefacts, to isolating or locking the device, terminating a process or blocking the action. FortiDLP also monitors the use of GenAI tools and SaaS applications, distinguishing between company and personal credentials, and reads sensitivity labels from Microsoft and Google Workspace.
FortiMail — the email channel. FortiMail has built-in email DLP with document fingerprinting (from Windows file shares or uploaded manually) and detection of sensitive data, such as financial, medical or personal data. Messages containing such data can be encrypted with IBE, which needs no software on the recipient’s side, and archiving with external storage makes it possible to reconstruct what went out and to whom. In API mode for Microsoft 365 and Google, FortiMail works without changing MX records and can claw back a message after it has been delivered.
How do we split the work in such a deployment? The tool provides the sensors and response mechanisms, but it does not decide which data is critical, which thresholds to adopt or who talks to the employee. We agree these with the client at the start: the list of critical data with its owners, the rules for each channel, the response ladder and the scope described in the work regulations. Signals from the data and email layers can be sent to the SOC, where they are correlated with other security events.
Where should an insider risk programme start in your organisation?
With a list of the data whose loss would hurt the organisation most, and with named owners for it — everything else follows from that list. An order that works in organisations of a few hundred people and up:
- Mandate and owner. The board appoints a programme owner, who sets up a team including HR, legal and security. Without a mandate the programme ends with a tool nobody operates.
- Critical data. A short list of data and systems with business owners. Five well-described items are worth more than fifty general ones.
- Map of access and channels. Who has access to the data on the list — including external, technical and privileged accounts — and which routes data can leave the company by.
- Legal basis and communication. The scope of monitoring described in the work regulations, a review with the legal team, information for employees before launch.
- Quick wins in access. Automatic HR notifications of departures and transfers, expiry dates for external accounts, a review of accounts without an owner.
- Rules for critical data. Warnings and observation first, to learn the norm; blocking only for channels with no business justification.
- Response ladder and a first exercise. Agreed thresholds, a decision path and a scenario rehearsed around the table before the first real case.
- Metrics and review. A first report for the board with time to remove access, channel coverage and the number of mistakes stopped.
Two elements of this plan depend more on people than on technology. Employee awareness — how to share files safely, use AI tools and whom to report a mistake to — is built in security awareness training, and how susceptible the team is to password phishing, where the compromised account scenario begins, is tested by phishing simulations. The specifics of the department that has access to every employee’s data are covered separately in our article on insider threats in HR.
If you want insider risk signals to be analysed around the clock together with other security events, see our SOC service; administrator and supplier accounts are brought under control as part of our PAM implementation. Let’s talk about which of these eight items is worth starting with in your organisation — contact us. More data protection material is available in the DLP and data protection hub.
Frequently asked questions
What is the difference between insider risk and insider threat?
Insider threat describes the danger itself — the possibility that a person with authorised access harms the organisation, deliberately or unintentionally. Insider risk is the way that danger is managed: deciding which data and which roles are most exposed, which signals reveal it and how the organisation responds to them. An insider risk programme therefore also covers mistakes and compromised accounts, not only malicious intent.
Is a DLP system enough to reduce insider risk?
No. DLP blocks or reports the transfer of data through a given channel, but it does not know that an employee has handed in their notice, moved to another department or holds a contractor account with no expiry date. HR and system owners have that knowledge. An effective programme combines technical signals with HR information in a single process, and DLP is one of its sensors.
Can an employer in Poland monitor business email and employee activity?
Yes, within the limits of Article 22³ of the Polish Labour Code: monitoring of business email and other forms of monitoring must serve the purposes set out in that provision, must not breach the confidentiality of correspondence or other personal rights of the employee, and their purposes, scope and method are set out in the work regulations, a collective agreement or a notice. Employees are informed no later than 2 weeks before monitoring starts. It is worth assessing a specific tool and its settings with the legal team before deployment.
How quickly should access be removed from an employee who is leaving?
The rules do not give a single deadline — control PS-4 in NIST SP 800-53 requires the organisation to define its own time period for disabling access and to make sure credentials are revoked. In practice the deadline depends on the risk of the role: privileged accounts and access to critical data are disabled the moment the engagement ends, and standard access no later than the leaving date. More important than the deadline is that HR triggers this path automatically, not on request.
Where should an insider risk management programme start?
With a list of the data and systems whose loss would hurt the organisation most, and with named owners for those assets. On that basis you establish who has access, which routes data can leave by and which signals will be observed. Tools are chosen only afterwards — for specific data and channels, not the other way round.
