Cybersecurity

Cybersecurity at the AI Inflection Point: A New Threat Architecture

By Clateway Technology Council · 15 March 2026 · 24 min read

CLATEWAY INSIGHTS · DEEP DIVE SERIES

Trend Report 2025–2026 · Part 2 of 6

Cybersecurity at the AI Inflection Point: A New Threat Architecture

How AI has permanently altered the cybersecurity landscape — transforming attack speed, collapsing the state-criminal boundary, and rendering the security playbooks of 2024 obsolete by 2026.

Based on 5,000 news signals, April 2025 – March 2026 | www.clateway.com

The second major trend from the 2025–2026 Clateway Media feed is one of the most consequential for enterprise IT leadership: cybersecurity has reached an inflection point driven by artificial intelligence that is not incremental, but structural. The threat landscape of 2026 is not a more intense version of 2024. It operates according to different mechanics, different actors, and different timelines — and the security architectures that most organisations built over the past decade are increasingly inadequate against it.

Four developments define this inflection point. First, the exploit-first cloud attack: Google's own threat intelligence confirmed in March 2026 that software vulnerability exploitation has overtaken credential theft as the primary initial access vector in cloud environments — a dramatic reversal of the pattern that shaped security investment for a decade. Second, AI-accelerated speed: CrowdStrike's 2026 Global Threat Report documented an average attacker breakout time of 29 minutes in 2025 — 65% faster than 2024, with some attacks completing lateral movement in 27 seconds. Third, the state-criminal convergence: Iran's Ministry of Intelligence and Security was documented directly operating through the commercial cybercriminal underground, using ransomware-as-a-service affiliates, commercial infostealers, and criminal botnets to achieve state objectives — while using 'hacktivist' front groups as cover. Fourth, the Stryker case study: a wiper attack on a $25 billion medical technology company, claimed by an Iranian-linked group, wiped an estimated 200,000 devices globally and disrupted EMS ECG systems in multiple US states — executed within days of a geopolitical escalation.

The leadership implication: Every enterprise security programme that has not been re-assessed against AI-enabled threats, exploit-first cloud attacks, and geopolitically motivated destructive malware is running on a threat model that is one to two years out of date. That gap is now measured in incidents, not risk ratings.

PART 1 THE EXPLOIT-FIRST CLOUD REALITY

The End of the Credential-First Era

A fundamental reversal in how cloud environments are compromised

For the better part of a decade, the dominant narrative in cloud security was built on a single premise: the most likely way an attacker would enter your cloud environment was through stolen or weak credentials, or misconfigured access controls. This narrative shaped an enormous investment cycle in identity and access management, MFA, privileged access management, and posture management tooling. It was largely accurate.

It is no longer accurate.

Google Cloud's H1 2026 Cloud Threat Horizons Report — published March 2026 and based on incident response and threat intelligence data from the second half of 2025 — documented a reversal that security leaders need to understand in precise terms: software vulnerability exploitation has overtaken credential abuse as the primary initial access vector in cloud environments for the first time since Google began publishing this report in 2021.

44.5% of cloud intrusions started with third-party software vulnerability exploitation in H2 2025

Source: Google Cloud Threat Horizons Report H1 2026

27% of cloud intrusions used credential compromise as primary entry vector — down from majority

Source: Google Cloud Threat Horizons Report H1 2026

2.9% was the share of exploit-first attacks in H1 2025 — by H2 2025 it had reached 44.5%

Source: Google Cloud Threat Horizons Report H1 2026

The numbers document a shift that happened within a single six-month period. In the first half of 2025, exploit-based initial access accounted for less than 3% of observed incidents. By the second half of the same year, it accounted for nearly half. This is not a gradual trend — it is an accelerating pivot that caught many security teams in mid-investment cycle.

Why the shift happened: AI-enabled exploitation

The explanation lies in the interaction between two simultaneous developments: improved credential security posture among cloud users, and AI-enabled acceleration of vulnerability exploitation.

Google's assessment is direct: the shift toward exploit-first attacks is likely partly a sign of defensive progress. Secure-by-default cloud configurations, improved credential hygiene, MFA enforcement, and identity monitoring have made credential-based initial access harder for attackers. Faced with better-defended credential surfaces, attackers shifted to the next available entry point: unpatched applications running on cloud infrastructure, particularly internet-facing services, virtual machines, and containers where the customer — not the hyperscaler — bears responsibility for patching.

The enabling technology is AI. As Saumitra Das of Qualys explained in analysis carried by Dark Reading: exploitation has become easier because of AI-driven vulnerability analysis, penetration testing, and exploit development. LLMs allow less technically sophisticated attackers to build well-crafted reconnaissance and exploitation frameworks, expanding the pool of actors who can execute sophisticated attacks. Critically, the time from vulnerability disclosure to mass exploitation has now collapsed from weeks to days.

The React2Shell case: 48 hours from disclosure to mass exploitation

The most vivid illustration of this new timeline came with CVE-2025-55182, the critical remote code execution vulnerability in React Server Components dubbed React2Shell. When the vulnerability was publicly disclosed in December 2025, Google Cloud observed multiple threat actors exploiting it to deploy XMRig cryptocurrency mining malware within approximately 48 hours of public disclosure. By January 2026, China-linked and North Korean nation-state actors had incorporated it into their attack chains.

The practical implication for IT leaders is stark. The era of the 30-day patching window, the 90-day patching window, or any patching window calibrated to historic exploitation timelines is over. In the AI-accelerated threat environment of 2026, the gap between vulnerability disclosure and active exploitation is measured in hours, not weeks.

Critical finding: Your patching cycle is now a competitive security differentiator. Organisations that can patch critical cloud application vulnerabilities within 24 to 48 hours of disclosure operate in a fundamentally different risk environment than those on monthly or quarterly cycles.

Identity still matters — but the model has changed

It is important to note that the exploit-first finding does not mean identity is irrelevant. Google's own analysis found that identity compromise underpinned 83% of compromises involving major cloud and SaaS-hosted environments — even when the initial access vector was software exploitation. The pattern is typically: exploit the unpatched application to gain a foothold, then steal or abuse credentials to move laterally through identity infrastructure.

Palo Alto Networks' Global Incident Response Report 2026 made a complementary finding: two-thirds of initial access was tied to identity in some way. The key change is that identity compromise now often follows software exploitation rather than leading it. The attack chain has changed even if the endpoint outcome has not.

Leadership implication: Security investment programmes that are solely identity-focused are now addressing the second step of the attack chain, not the first. Cloud application patch management has become a frontline security control — not an IT operations convenience.

PART 2 AI AS ATTACKER: SPEED, SCALE, AND THE ARMS RACE

The AI Arms Race: 29 Minutes, 27 Seconds, and What It Means

The CrowdStrike Global Threat Report: 2025 in numbers

CrowdStrike's 2026 Global Threat Report — drawing on data from more than 280 named adversaries across 2025 — defined the year as 'the year of the evasive adversary'. The headline finding was one of the starkest metrics in recent cybersecurity history: the average attacker breakout time — the period between initial access and lateral movement to other systems — fell to 29 minutes in 2025. This represents a 65% acceleration from 48 minutes in 2024, and a dramatic compression from 98 minutes in 2021. The fastest observed breakout time on record: 27 seconds. In one documented intrusion, data exfiltration had begun four minutes after initial access.

29 min average eCrime breakout time in 2025 — 65% faster than 2024

Source: CrowdStrike 2026 Global Threat Report

27 sec fastest observed breakout time ever recorded

Source: CrowdStrike 2026 Global Threat Report

89% increase in attacks by AI-enabled adversaries in 2025 vs. 2024

Source: CrowdStrike 2026 Global Threat Report

82% of threat detections in 2025 were malware-free — intrusions moved through trusted pathways

Source: CrowdStrike 2026 Global Threat Report

These numbers demand a specific leadership response, because they describe a threat environment that manual incident response processes cannot operate within. The classic security operation model — detect, investigate, escalate, respond — was designed for environments where attackers moved at human speed and needed days or weeks to complete an intrusion. An adversary who achieves lateral movement in 29 minutes on average, and in 27 seconds in the most capable cases, has already completed the intrusion before most security teams have acknowledged the first alert.

How AI is being used by threat actors in 2025–2026

The CrowdStrike data documents five specific ways in which AI is being integrated into attacker operations, each with distinct implications for defenders:

Reconnaissance acceleration

AI-powered reconnaissance allows threat actors to map an organisation's attack surface, identify unpatched vulnerabilities, and profile employee targets faster and at greater scale than human-led research. In particular, the Clateway feed documented cases of attackers using LLMs to automate the identification of exposed cloud services, misconfigured APIs, and vulnerable application versions — compressing what previously required hours of manual research into minutes.

AI-generated phishing and social engineering

The most thoroughly documented AI application in attack operations is phishing content generation. Research cited in the Clateway feed documented LLM-generated spear phishing achieving a 54% click-through rate, compared to 12% for human-generated phishing — a 4.5-fold improvement in effectiveness. Separately, analysis cited in the feed found that agentic AI has improved spear phishing effectiveness by 55% since 2023. The Darcula phishing-as-a-service platform received an AI upgrade in 2025 that made sophisticated phishing attacks accessible to technically unsophisticated attackers. Voice phishing (vishing) — which saw a 442% increase between the first and second halves of 2024 — continued to grow through 2025, with AI-powered voice cloning enabling convincing impersonation of internal IT staff, executives, and support personnel.

AI-generated malware

VoidLink — documented in January 2026 as a modular Linux malware framework — was found to have been built almost entirely by AI agents, marking a significant milestone in autonomous malware development. AI-generated Linux miner 'Koske', documented in July 2025, was found to include features that met or exceeded what traditional human-developed malware could produce. Pakistan's APT36 threat group was documented using vibe-coding tools to generate malware at scale, even if individual samples were of lower quality — the scale advantage compensating for quality deficiencies. Russia's Fancy Bear (FANCY BEAR) deployed LLM-enabled malware named LAMEHUG, embedding prompt-based logic to automate reconnaissance and document collection.

Defence evasion and forensic erasure

AI is being used to make attacks harder to detect and investigate. Punk Spider (behind Akira ransomware) was documented using AI-generated scripts to accelerate credential dumping and systematically erase forensic evidence. Separately, 82% of CrowdStrike's detections in 2025 were malware-free — attackers moved through legitimate credentials, trusted identity flows, and approved SaaS integrations, deliberately avoiding the IOC footprints that traditional security tools detect. When attacks don't involve malware, legacy detection systems produce no signal.

AI systems as attack targets

A category of attack that was essentially non-existent in 2023 has become material by 2026: attacks that target AI systems themselves. CrowdStrike responded to incidents at more than 90 organisations where adversaries injected malicious prompts into legitimate generative AI tools deployed in enterprise operations. ClickFix attacks leveraged AI-generated content summaries to push malware. The 'ShadowLeak' attack exploited ChatGPT infrastructure to silently exfiltrate company data with no trace on enterprise systems. As AI tools become embedded in enterprise workflows, they become attack vectors as well as productivity platforms.

The defining metric: If your mean time to detect is measured in hours and your mean time to respond is measured in days, you are operating in a threat environment where the average attacker has already achieved their objective and moved on before you have identified the incident.

Leadership implication: Speed is now the primary competitive variable in cybersecurity. Organisations cannot out-skill adversaries who operate at machine speed — they must out-automate them. Detection, decision, and response must be automated to the greatest extent possible, with humans reserved for decisions that require judgment.

PART 3 STATE-CRIMINAL CONVERGENCE

The State-Criminal Convergence: Iran's MOIS and the Collapse of Attribution

A new threat category that security operations teams are not equipped for

One of the most strategically significant shifts documented in the Clateway feed from 2025 through March 2026 is the deepening integration between state intelligence services and the commercial cybercriminal underground. This is not new in concept — Russia has long employed civilian hackers and organised crime networks for state operations. What is new is the sophistication and operational depth of the integration, particularly for Iran.

Check Point Research, publishing in March 2026, documented what they described as a fundamental shift in Iranian cyber operations: Iran's Ministry of Intelligence and Security (MOIS) has moved from merely impersonating cybercriminals to directly integrating criminal tools, infrastructure, and even affiliate relationships into state-directed operations.

How MOIS has embedded in the cybercriminal ecosystem

The Check Point analysis, corroborated by Palo Alto Networks Unit 42 and other sources in the Clateway feed, documented four specific mechanisms of state-criminal integration:

Criminal malware in state operations

Void Manticore — the MOIS-linked APT group that operates the 'Handala' hacktivist persona — was documented using the commercial infostealer Rhadamanthys as a core element of its attack chains. Rhadamanthys is sold on darknet forums and is used by both criminal actors and state-aligned groups. By adopting commercial criminal tools, MOIS operations become harder to attribute and attribute correctly — a SOC that sees Rhadamanthys activity may classify it as criminal rather than state-sponsored.

RaaS affiliate participation

The most striking example of state-criminal convergence in the feed period was the October 2025 attack on Israel's Shamir Medical Center. The attack was initially claimed by the Qilin ransomware group and attributed to Eastern European criminals. Three weeks later, Israel's National Cyber Directorate corrected the attribution, identifying Iranian state-affiliated operators as having acted as ransomware-as-a-service affiliates. The Iranian state was using criminal ransomware infrastructure to achieve strategic objectives while maintaining plausible deniability through the RaaS brand. As Sergey Shykevich of Check Point stated: 'there can be a case where a SOC or CISO will see something in their network that they associate with cybercrime activity and label it as low risk. And in reality, it will be an Iranian threat actor who will be able to execute destructive activities.'

Shared criminal infrastructure

MuddyWater — the MOIS-linked group responsible for widespread espionage operations across the Middle East, Africa, and Europe — was found to have infrastructure and code-signing certificate overlaps with the Tsundere/DinDoor botnet, a criminal botnet that uses Node.js and Deno-based execution methods. The botnet has been linked to MuddyWater through multiple independent sources. This represents not just tool sharing but operational infrastructure sharing — the state and criminal ecosystems are using the same command and control nodes.

Coordinated multi-actor operations

CSIS analysis of over 250,000 Telegram messages from more than 178 hacktivist and proxy groups revealed that during the June 2025 Israel-Iran military conflict, the timing, target selection, and exchange of vulnerabilities and attack scripts across ostensibly independent hacktivist groups showed patterns of orchestration rather than organic activity. The analysis concluded that these campaigns indicated institutional backing and coordination rather than spontaneous hacktivist mobilisation. Multiple 'independent' hacktivist groups were effectively extensions of state cyber operations.

The Stryker incident: a case study in geopolitically motivated destructive attack

On the morning of March 11, 2026 — two weeks after US-Israeli military operations known as 'Operation Epic Fury' began against Iran — Stryker Corporation, one of the world's largest medical technology companies with over $25 billion in annual revenue, suffered a coordinated wiper attack. By mid-morning, employees across 79 countries could not access their systems. Mobile devices had been wiped to factory settings. Login screens on some devices displayed the logo of 'Handala' — the ostensibly pro-Palestinian hacktivist group that is, in fact, Void Manticore, a MOIS-run APT.

The scale of disruption was extraordinary by enterprise standards:

• Stryker internally described the attack as a 'severe, global outage impacting all Stryker laptops and systems that connect to our network.'

• Handala claimed 200,000 systems, servers, and mobile devices wiped; 50TB of data exfiltrated; operations disrupted in 79 countries.

• Maryland's Institute for Emergency Medical Services notified hospitals statewide that Stryker's Lifenet ECG transmission system was 'non-functional in most parts of the state,' requiring EMS clinicians to use radio consultation rather than electronic transmission.

• Ireland's National Cyber Security Centre launched an investigation. Microsoft engineers assisted with forensic analysis and recovery.

• Stryker shares fell more than 3% following media reports of the breach.

• The Stryker attack represented the first time Handala/Void Manticore had targeted a major US business — a significant escalation noted by Check Point Research.

The attack was enabled by a technique that Iran has used before and that Tenable's analysis documented as increasingly prevalent: pre-positioning. Nation-state hackers often infiltrate an organisation months before any visible action. The Stryker attack likely involved access gained before February 28, 2026 — the attackers were in position and waiting for a geopolitical trigger. When US-Israeli strikes on Iran began, the command was executed.

Iran concurrently announced new targeting priorities, specifically naming Amazon AWS, Google, Microsoft, Oracle, Palantir, and Nvidia as organisations that the IRGC had threatened to attack. Drone strikes had already disrupted AWS data centres in the UAE and Bahrain. The Stryker incident was not an isolated event — it was the first visible operation of a broader retaliatory campaign.

Board-level concern: The Stryker incident demonstrates that enterprise IT infrastructure is now a legitimate target in geopolitical conflicts, even for organisations with no direct government or defence connections. The trigger for a destructive attack can be a missile strike in a country your company has no operations in.

Leadership implication: The state-criminal convergence means your SOC's attribution logic is now a security control. An incident classified as 'low-risk criminal activity' may be a MOIS or FSB operation. Update your threat classification model to treat unattributed ransomware and infostealer activity at critical targets as potentially state-sponsored until demonstrated otherwise.

PART 4 THE AI-POWERED PHISHING EPIDEMIC

Phishing Reinvented: When AI Breaks the Human Firewall

The effectiveness gap that training cannot close

For years, the standard enterprise response to phishing was a combination of email gateway filtering and employee awareness training. Both were imperfect, but together they created sufficient friction to limit the volume of successful attacks. The AI-powered evolution of phishing in 2025–2026 has fundamentally altered that equation — and the alteration is not marginal.

The data points are consistent across multiple sources in the Clateway feed. LLM-generated spear phishing achieves a 54% click-through rate versus 12% for human-generated attempts. AI-assisted spear phishing is 55% more effective than previous techniques. AI-generated phishing content personalised to individual targets is now being produced at scale — the Darcula PhaaS platform's AI upgrade in 2025 made it accessible to non-technical actors. Traditional awareness training was designed to help employees recognise generic phishing lures. It was not designed for, and cannot reliably defend against, personalised, contextually accurate, grammatically flawless attacks generated by LLMs with access to LinkedIn, public company information, and email address formats.

The evolution of attack vectors: phishing is going mobile

The Clateway feed documented a consistent theme across 2025–2026: phishing is migrating away from email — where defences have improved — toward channels where enterprise defences are structurally weaker. The most significant developments:

Telephone-oriented attack delivery (TOAD)

TOAD attacks — where the only payload in a phishing email is a phone number — are bypassing email gateways entirely, because there is nothing malicious in the message for content filters to detect. Victims call a number and are socially engineered by a human operator or, increasingly, an AI voice agent impersonating IT support or a vendor. The callback rate for TOAD campaigns is significantly higher than for link-based phishing, partly because the phone call triggers social compliance norms that email does not.

AI voice cloning and vishing

Voice phishing (vishing) increased by 442% between the first and second halves of 2024, continuing its growth through 2025. AI-powered voice cloning has eliminated one of the traditional tells of vishing attacks — an unfamiliar accent or unnatural speech pattern. A researcher-developed framework documented in September 2025 demonstrated real-time voice cloning that could impersonate specific individuals during live phone calls, with sufficient fidelity to compromise help desk workflows that rely on voice authentication. ShinyHunters' expansion into vishing for SaaS data theft — documented in January 2026 — showed this technique being operationalised at scale by financially motivated criminal groups.

QR code phishing (quishing)

QR codes have become a common phishing vector because they bypass email link filtering — email security tools cannot easily assess where a QR code leads. The Anti-Phishing Working Group documented a significant increase in QR-code-based phishing campaigns through 2025. North Korean Kimsuky group was specifically cited by the FBI for quishing attacks against US and foreign government agencies.

SMS and mobile-first attacks

Chinese 'Smishing Triad' groups — which became notorious through fake toll and package tracking texts — were documented shifting toward lower-frequency, higher-impact government impersonation attacks. 'Lighthouse', a phishing-as-a-service operation Google worked to disrupt in November 2025, had powered massive volumes of mobile phishing campaigns across the US. The shift is structural: mobile devices have weaker phishing defences, users have higher trust in SMS communications, and enterprise mobile security controls are less mature than email security.

When phishing breaks the AI itself

A final category documented extensively in the Clateway feed is the use of phishing-style techniques to compromise AI systems deployed in enterprise operations. The 'ShadowLeak' attack in September 2025 exploited ChatGPT infrastructure to exfiltrate company data without leaving traces on enterprise systems. EchoLeak — a zero-click Microsoft Copilot vulnerability disclosed in June 2025 — could have enabled data exfiltration via AI-generated summaries triggered by malicious content in emails or documents. ClickFix attacks evolved to include AI-generated content summaries that pushed malware installation — victims were more likely to follow instructions that appeared to come from an AI tool they trusted.

The pattern is consistent: as AI tools become integrated into enterprise workflows, the attack surface for social engineering expands. Attackers are now crafting content not to deceive humans, but to deceive or exploit the AI systems that humans have delegated decisions to.

Leadership implication: Employee phishing awareness training has not lost its value, but it has lost its sufficiency. Technical controls — identity verification for help desk calls, hardware MFA resistant to real-time phishing, AI-aware email and mobile security, and explicit policies about what actions AI assistants are authorised to take — must carry the weight that human judgment alone cannot bear.

PART 5 THE DISASTER RECOVERY REALITY

Disaster Recovery in the Age of Wiper Attacks

Why the Stryker incident is a stress test most organisations will fail

The Stryker attack illuminated a gap in enterprise business continuity and disaster recovery planning that Dark Reading's post-incident analysis identified directly: most DR plans are designed for infrastructure failures, not adversarial destruction. They test against scenarios where systems go offline due to hardware failure, software bugs, or ransomware encryption. They rarely test against scenarios where an adversary with Active Directory access wipes all endpoints simultaneously, in 79 countries, within minutes.

In Stryker's case, security experts assessed that the attackers gained access to Stryker's Active Directory services and used Microsoft Intune — the legitimate enterprise endpoint management platform — to issue a device wipe command at scale. The attack did not require the deployment of custom malware to each endpoint. It leveraged the organisation's own management infrastructure against itself. This is a qualitatively different kind of attack than the ransomware scenarios that most DR tabletops simulate.

The three DR gaps the Stryker incident exposed

Gap 1: Active Directory and identity infrastructure as a single point of failure

If an attacker gains control of Active Directory, they have effective control of every endpoint, server, and service connected to it. Most enterprise DR plans do not include a scenario where Active Directory itself is compromised. The recovery from AD compromise — rebuilding directory services, re-issuing credentials, re-enrolling devices — can take weeks, not hours, and requires a level of preparation that most organisations have not made.

Gap 2: Endpoint management platforms as attack amplifiers

Microsoft Intune, Jamf, and equivalent endpoint management platforms are powerful administrative tools. They are also powerful attack amplifiers if an attacker gains sufficient privileges. An attacker who can issue commands through Intune can wipe every managed device in the organisation simultaneously. Most DR plans treat endpoint management tools as part of the recovery toolkit. They should also be treated as a risk that requires protective controls — including multi-person authorisation for mass device operations, network segmentation of management infrastructure, and out-of-band verification for destructive commands.

Gap 3: Operational continuity without IT

For a medical technology company like Stryker, the DR failure was not abstract. Maryland hospitals lost access to ECG transmission systems that paramedics relied upon for real-time cardiac monitoring during transport. The medical impact was limited in this case, but the scenario — enterprise IT failure cascading into critical operational system failure — is one that every organisation in healthcare, manufacturing, financial services, and critical infrastructure needs to have tested.

What a resilience programme must include in 2026

The Clateway feed's security signals from 2025–2026 point to five requirements for a DR programme that is fit for the current threat environment:

• Adversarial DR testing: Run tabletop exercises against a sophisticated adversary who has gained Active Directory access — not a ransomware scenario alone. Measure recovery time from AD compromise, not just from data encryption.

• Out-of-band management: Ensure that critical infrastructure can be managed and communicated about through channels that are not dependent on the corporate Microsoft environment. This includes emergency communication plans that do not rely on corporate email, Slack, or Teams.

• Endpoint backup and re-provisioning: Maintain clean device images, secure out-of-band backup systems, and tested processes for emergency re-provisioning of endpoints. The ability to rebuild a device estate at speed is now a DR capability requirement.

• Privileged action controls: Implement multi-person authorisation for any mass device management operations, configuration changes to identity systems, and mass account modifications.

• Cascade impact mapping: For each critical IT system, map the operational systems that depend on it. Know in advance what happens when Intune goes dark, when AD is compromised, when your SaaS CRM is inaccessible. Have manual fallback procedures documented and tested.

The CrowdStrike benchmark: One year after the July 2024 CrowdStrike outage — which affected 8.5 million Windows endpoints — the post-incident review found that organisations had not learned the lesson about technology concentration risk. The Stryker wiper attack is the adversarial version of that event. The lesson is the same: concentration of critical operational functions in a single IT dependency is an existential risk.

PART 6 THE LEADERSHIP AGENDA

The RCEC Response: Security in the AI-Inflection Era

The four developments documented in this article — exploit-first cloud attacks, AI-accelerated attacker speed, state-criminal convergence, and AI-powered phishing — are not independent trends. They are mutually reinforcing dimensions of a single structural shift: cybersecurity has entered an era in which the primary variable is not the sophistication of any individual attack, but the fundamental speed advantage that AI-enabled adversaries hold over manually-operated defences.

Responding to this shift requires decisions at every level of the RCEC framework:

Research: Intelligence for a new threat model

• Rebuild your threat model: Explicitly include AI-enabled attack vectors, exploit-first cloud intrusion patterns, state-criminal actor scenarios, and destructive wiper attacks. A threat model that was built in 2023 or 2024 is operating on assumptions that no longer hold.

• Instrument for exploit detection: If your primary detection logic is built around malware signatures and known-bad indicators, you are detecting 18% of modern intrusions. Behaviour-based detection, anomaly detection in identity flows, and application-layer monitoring for exploitation patterns are now baseline requirements.

• Subscribe to operational threat intelligence: The geopolitical context of your threat environment now matters. The Stryker attack was preceded by weeks of public warnings about Iranian cyber retaliation. Organisations that were monitoring CISA advisories and commercial threat intelligence had warning. Many were not watching.

Commit: Board-level security decisions

• Authorise automated response: The 29-minute average breakout time makes manual incident response inadequate as a primary defence. Authorise security operations to implement automated detection and containment capabilities — including automated network isolation, automated credential revocation, and automated alert escalation — without requiring manual approval for each action.

• Reframe patching as a security control: The collapse of the vulnerability-to-exploitation window makes patching velocity a board-level metric. Set a target for critical vulnerability patching within 48 hours of public disclosure for internet-facing cloud applications. Report this metric to the board alongside other security KPIs.

• Commission an AD resilience review: Assess your organisation's ability to detect, contain, and recover from an Active Directory compromise. If you cannot answer the question 'how long would it take to rebuild our AD from scratch?' with confidence, you have a material DR gap.

Execute: The operational capabilities

• Deploy AI-powered defence: Adversaries are using AI to operate at machine speed. Defenders must do the same. AI-powered security operations platforms — including SIEM with AI-driven detection, automated threat hunting, and AI-assisted incident response — are now operational necessities, not advanced capabilities.

• Harden phishing-resistant MFA: FIDO2 hardware keys or passkeys eliminate the real-time phishing vulnerability of TOTP-based MFA. Given the documented effectiveness of AI-powered phishing and vishing, organisations in high-risk sectors should be prioritising phishing-resistant MFA for all privileged accounts as a matter of urgency.

• Establish out-of-band communication: Define and test an emergency communication process that is completely independent of your corporate Microsoft environment. This includes a secondary channel for IT staff coordination, a mechanism for customer and partner notification, and a documented decision authority structure that does not require access to corporate systems.

Control: Monitoring for the AI-era threat

• Measure MTTD and MTTR against the 29-minute standard: If your mean time to detect exceeds 29 minutes for a cloud intrusion, the average attacker has already achieved lateral movement before detection. Benchmark your current MTTD and MTTR, identify the gap, and build a roadmap to close it.

• Monitor endpoint management platforms: Intune, Jamf, and equivalent platforms should be instrumented with anomaly detection for high-impact operations — particularly mass device commands, policy changes, and bulk account modifications. Alert on any operation that affects more than a defined threshold of endpoints in a single action.

• Include geopolitical triggers in your threat monitoring: Subscribe to threat intelligence that covers geopolitical context. When nation-state cyber escalation is publicly anticipated — as it was before the Stryker attack — elevate your alert posture, review exposure of critical systems, and activate pre-positioned response capabilities.

Conclusion: Security Has Crossed a Threshold

The cybersecurity signals in the Clateway Media feed from April 2025 to March 2026 do not describe an incremental worsening of a familiar threat landscape. They describe a threshold crossing. The combination of AI-accelerated attack speed, the exploit-first reversal in cloud intrusions, the operational integration of state and criminal actors, and the demonstrated effectiveness of AI-powered phishing against the human firewall represents a qualitative change in the nature of the threat.

The organisations that will navigate this environment effectively are not necessarily those with the largest security budgets or the most sophisticated tools. They are the organisations that have made the governance decisions — about automation authority, patching velocity, DR investment, and threat intelligence consumption — that allow them to operate at a speed appropriate to the threat. Speed, not complexity, is the defining variable.

For Clateway clients and the IT leaders who read Clateway Insights, the message is direct: the security architecture you built for a 2023 threat environment is operating at a structural disadvantage in 2026. The gap is not insurmountable, but it requires deliberate action — and the window for deliberate, unhurried remediation is shorter than it was twelve months ago.

The adversary is operating at machine speed. Your defence must too.

Key Statistics Referenced in This Article

Cloud attacks: Software exploitation overtook credentials as primary cloud initial access vector in H2 2025. Third-party software entry: 44.5% of incidents (from 2.9% in H1 2025). Credential abuse: 27%. Window from vulnerability disclosure to mass exploitation: now days, not weeks (Google Cloud Threat Horizons H1 2026).

Attack speed: Average eCrime breakout time: 29 minutes in 2025 (down from 48 in 2024, 98 in 2021). Fastest observed: 27 seconds. 82% of detections malware-free. 89% increase in AI-enabled attacker operations. 37% increase in cloud-conscious intrusions (CrowdStrike 2026 Global Threat Report).

Phishing effectiveness: LLM-generated spear phishing click-through rate: 54% vs 12% for human-crafted. AI-assisted spear phishing 55% more effective since 2023. Vishing increased 442% in H2 2024 vs H1 2024.

Stryker/Iran: 200,000 systems claimed wiped. 50TB data exfiltrated claimed. Operations disrupted in 79 countries. ECG systems non-functional in most of Maryland (March 11, 2026). Handala = Void Manticore = MOIS APT (Check Point, Palo Alto Networks Unit 42, CrowdStrike).

State-criminal convergence: Iran MOIS using commercial infostealers, RaaS affiliates, criminal botnets in state operations. Shamir Medical Center attack initially attributed to Qilin/Eastern European crime — corrected to Iranian state operators. North Korea FAMOUS CHOLLIMA's PRESSURE CHOLLIMA stole $1.46B in crypto — largest single financial theft ever reported.

About Clateway

Clateway is a fractional executive leadership practice serving IT leaders across regulated industries in Europe. Through the RCEC methodology — Research, Commit, Execute, Control — Clateway helps organisations make better technology decisions, faster. This article is Part 2 of the Clateway Insights Trend Report 2025–2026, covering the six dominant trends identified from 5,000 news signals in the Clateway Media feed. For advisory engagements, executive roundtables, or the Clateway Academy fractional leadership programme, visit www.clateway.com.

Topics: cybersecurity

Related articles

• AI and Cybersecurity: The Growing Threat of AI-Powered Identity Fraud – Artificial intelligence is revolutionizing identity fraud, enabling sophisticated attacks through deepfakes, voice cloning, and automated social engineering. This article explores the evolving…
• AI Agent Security: The New Identity Management Frontier Demanding C-Suite Attention – As AI agents proliferate across enterprise systems with unprecedented autonomy and access privileges, organizations face a fundamental shift in identity and access management. The emergence of…

How Clateway can help

• Technology and cyber board advisory – Independent technology and cyber expertise for boards and supervisory boards: advice on management's plans, NIS2 accountability, AI governance and director…
• NIS2 compliance and readiness – Get NIS2 compliant: readiness assessment, AI-generated NIS2 policies, risk register, supplier security, incident management and board reporting in one platform.
• Vulnerability management – Scan your systems and technologies for CVEs, prioritise with EPSS and CISA KEV, track SLAs, sync Dependabot alerts and watch SaaS vendor advisories in one…