ENISA Threat Landscape 2026: What 8,257 Incidents Tell Mid-Market Security Leaders

ENISA’s Threat Landscape 2026 landed on 22 September 2026, and its headline number invites the wrong reading. The agency collected and analysed 8,257 incidents recorded across calendar-year 2025, drawn mainly from open sources plus anonymised information shared voluntarily by EU Member States and members of the ENISA Cyber Partnership Programme. ENISA is explicit that these inputs “do not constitute a complete picture of the cyber threat landscape”. Treat 8,257 as a structured sample of what gets reported, not a census.

Two methodology changes reinforce the caution. The reporting window moved to the calendar year, overlapping six months with the previous edition, and ENISA expanded the cybercrime activities it tracks, adding data breaches and fraud. Do not compare editions and call the movement a trend, though ENISA notes the expansion did not substantially change the rankings. Expect the next edition around September 2027. Until then this is the reference document for 2026 risk planning.

Secondary coverage stops at the distribution chart. The useful content sits one level below it: ENISA’s own distinction between what got counted and what mattered.

Volume is not impact

The reading that matters is volume against impact. DDoS attacks account for 51.3% of recorded incidents, a distribution ENISA describes as dominated by low-impact DDoS attacks; unauthorised access follows at 39.5%, dominated by financially motivated activity. ENISA does not pretend the counts tell the whole story. In its NIS Investments report, representatives of EU organisations in highly critical NIS sectors characterised DDoS as “noise”, while ransomware dominated their concerns. The executive summary draws the same line: “Financially motivated activities (29.3% of all recorded incidents), particularly ransomware, remained the most impactful incidents in the short-term.”

Much of the raw volume is ideology-driven: by assessed objective, ideology-driven operations represent 57.3% of observed activity, mostly hacktivist-led DDoS campaigns against public-facing services, and ENISA notes these claims “did not result in large-scale or significant impact” even where they dominate the charts. State-nexus cyberespionage sits at just 5.9%, barely a mark on a volume chart and enormously consequential over multiple years.

That contrast is the report’s most useful lesson for anyone accountable for a security program: the activity that fills alert queues is not the activity that ends up in front of a board, an auditor or a customer. Teams that measure themselves in incident counts will keep proving they were busy; teams that measure themselves against materialised impact first have to decide what impact means for their business.

The incidents that move numbers

Ransomware covered 47.3% of financially motivated activity in 2025. One measurement caveat applies: ENISA largely measures visibility through claims posted on data leak sites, so the ranking is a proxy for victim counts; companies that pay and move on never appear there. With that understood, manufacturing carried 25.2% of claims, business services 18.7%, a catch-all ‘other’ category 10.7% and public administration 6.7%.

How intruders get in matters as much as what they do inside. Exploiting N-day and 0-day vulnerabilities remained, in ENISA’s words, “a prevalent intrusion vector to enable unauthorised access (60.4%)”; that share is drawn from unauthorised-access incidents where ENISA could identify an intrusion vector, a small subset of the dataset. The direction still matches every patching backlog a mid-market team owns: internet-facing systems missing known fixes remain the cheapest way in, and the cheapest way in gets used. Product vendors should also note the regulatory clock; the Cyber Resilience Act puts actively exploited vulnerabilities on mandatory reporting timelines (see CRA 24-hour vulnerability reporting).

Social engineering is the other half of the intrusion path. Among the incidents where social engineering techniques were identified, phishing accounted for 77.8%, increasingly industrialised through phishing kits and phishing-as-a-service platforms; the ClickFix technique, which talks a user into running attacker commands as if they were fixing a problem, and smishing both rose. None of this is technically novel. It persists because it works, and it works at scale.

Dependencies are the sleeper item. ENISA flags that supply chains, third-party providers, cloud environments and other cyber dependencies “continued resulting in impactful and large-scale incidents”, citing compromises of popular libraries and npm packages, among them the Shai-Hulud campaign. ENISA published a technical advisory on package-manager security in March 2026. When one compromised dependency reaches thousands of downstream companies overnight, dependency management stops being a hygiene topic and becomes incident-response planning.

Ransomware at the top of the impact tier is also why insurability belongs in the same conversation. Underwriters now probe the specifics behind a renewal: identity paths, tested recovery, third-party visibility. If a renewal lands in 2026, ENISA’s impact-tier numbers are a reasonable basis for judging readiness (see cyber insurance renewal readiness).

Where finance actually sits

Sector rankings need careful handling: two denominators look similar and mean different things. Across all recorded incidents, public administration was the most impacted sector at 31.8%, followed by business services (8.5%), transport (8%), manufacturing (6.9%) and finance/banking (5.6%). Much of the public-administration share is ideology-driven DDoS, the category already deflated in the volume discussion. ENISA is direct about that: the public-administration picture is “largely impacted by ideology-driven DDoS attacks”.

If you run a fintech, the cybercrime breakdown is the more relevant table. Across all cybercrime incidents, manufacturing was the most impacted sector at 14.4%, followed by business services (13.5%), the catch-all ‘other’ category (11%) and finance/banking (9.2%). ENISA’s explanation is blunt: financially motivated operators primarily target sectors handling monetisable data, and finance appears in the top ranks of both lists. The correct conclusion is not that fintech is a low-priority target; it is that in the table organised around money, finance sits fourth rather than first, and the incidents most likely to touch a fintech’s numbers are not the ones dominating headline volume.

Sector tables describe where incidents were observed, not the odds that a specific company gets hit. Your exposure lives in your own attack surface, your dependencies and the resale value of your data; use ENISA’s tables to frame the conversation, then answer it with your own environment.

AI in the playbooks, and on the attack surface

The AI chapter deserves a slow read. Its long view is both sober and unsettling: “During 2025, ENISA observed AI being incorporated into the playbooks of cybercriminal, State-nexus and information manipulation sets, although often in a supporting role rather than as a primary capability. This is highly likely to change as so-called ‘frontier models’ would become more widely available.”

The 2025 reality was less cinematic than the coverage suggested. ENISA reports attackers “primarily use consumer-grade AI tools to augment existing skills and adapt attack vectors rather than to achieve breakthrough capabilities”. The concrete cases still earn attention: the Gentlemen ransomware group used AI coding assistants to build its ransomware-as-a-service panel, and a fake Claude website distributed malware. Supporting role, real output.

The supporting-role finding is worth sitting with, because it argues against both extremes: against the brief that autonomous AI attackers are already inside everyone’s perimeter, and against treating AI-enabled attacks as a problem 2026 planning can defer. In 2025, AI showed up where it produced immediate value for adversaries. Assume the frontier-model shift arrives the same way, in whichever phase of the attack chain saves the adversary the most time.

The forward view carries the operational weight. ENISA assesses that while AI will “highly likely continue to primarily act as a force multiplier”, it is “also likely 2026 will see an increased number of the kill chain’s phases being directly enabled by AI, with possible experimentation of Human-out-of-the loop proof of concepts, which would be likely to hinder detection and imputation for networks defenders”. That is assessment language, an outlook rather than an observation. The operational consequence matters: if attack chains compress, the window between initial access and impact shrinks, and detection assumptions built around human-speed tradecraft age first.

Then there is the side of the story that has nothing to do with attackers. “The integration of AI systems into enterprise environments creates a new attack surface.” ENISA adds that AI applications “are increasingly becoming targets, particularly where they have access to files, credentials, browser sessions or development environments”. Every AI tool adopted since 2023 is effectively a new third party inside the system boundary, holding tokens that outlive whoever created it. The regulatory direction matches: the EU’s Action Plan on Cybersecurity and AI builds on the AI Act, the Cyber Resilience Act and NIS 2, and AI governance expectations are moving from principles toward evidence.

What I would put on the 2026 list

First, impact-weighted monitoring. Most mid-market teams do not lack telemetry; they lack a decision about which alerts represent business risk. The weighting argument is already in ENISA’s numbers: if 51.3% of observed incidents are DDoS and NIS-critical organisations call that noise, DDoS belongs in automated edge-level mitigation, not analyst queues. Depth belongs to the classes that carry impact: unauthorised access, its precursors (exploited vulnerabilities, credential capture through phishing) and dependency anomalies. Playbooks should be sized to extortion and third-party scenarios, not traffic spikes; a tabletop that rehearses a compromised vendor will do more for the 2026 program than a dashboard that counts volumetric attacks.

Second, dependency and third-party governance, extended to AI dependencies. The large-scale incidents ENISA highlights reached victims through vendors, libraries and cloud platforms their own staff did not control. Map what the business runs on, then decide which dependencies can be severed for a week without halting operations. The vendor inventory is rarely the difficult part; knowing what the business can survive without is. The banking-sector TPRM overhaul is a readable template for the scrutiny now heading toward every sector (bank TPRM guidance overhaul), and AI services belong on the same register, under the same questions.

Third, AI-tool attack surface controls. ENISA’s observation that threat actors treat AI systems both as tools and as targets has a practical edge: the question is what your AI applications can reach. Inventory them the way ENISA’s targeting pattern implies: which applications read files, touch credentials, drive browser sessions or reach development environments, and under whose identities do they act? Route that access through scoped service identities instead of standing personal tokens, require human approval for actions that write, send or spend, and log enough for an investigation to reconstruct. The application-level control layer is covered in our OWASP LLM Top 10 breakdown, and for agentic tools chained to files, credentials and browsers, MCP security design considerations covers the decisions most teams have not made yet.

None of this demands a bigger security budget. It demands a sharper one, pointed at the split ENISA documents so plainly: the incidents that dominate the counts are mostly noise, and the ones that dominate impact arrive through unpatched attack surface, captured credentials and dependencies nobody fully owns. As of the Threat Landscape 2026 data, that is where the money sits, and where 2026 planning should start.