Tomislav Vazdar

The Digital Omnibus, Regulation (EU) 2026/1744, entered into force on 27 July 2026 and moved Annex III high-risk duties from 2 August 2026 to 2 December 2027, with embedded high-risk systems in regulated products moving to 2 August 2028.

• Article 50 did not move: transparency obligations have applied since 2 August 2026.
• Conversational systems must tell people they are talking to a machine, synthetic content must be marked, emotion recognition and biometric categorisation require disclosure, and deepfakes must be labelled — none of this depends on whether a system is high risk.
• Systems already on the market have until 2 December 2026 to implement machine-readable marking under Article 50(2).
• The obligation to conduct a conformity assessment and maintain a quality-management system and technical documentation was pushed back.
• The obligation to know how many generative tools are running inside an organisation is live now — and most organisations cannot answer it.

The post closes by asking readers whether they can account for every generative tool in use: "not the sanctioned ones. All of them."

In a later post, they call the 2 December 2026 marking deadline the first real test of whether the transparency rules have teeth. Unlike the high-risk regime, it is easy to check from the outside: "Anybody can look."

Read original

They watch Croatia's national CSIRT quarterly review for two numbers: whether the count of categorised entities keeps growing, and whether incident profiles shift away from known vulnerabilities and unprotected remote access.

• If they do not, the basics still are not landing two years in.
• On the Collins Aerospace MUSE attack that disrupted check-in at Heathrow, Brussels and Berlin, they separate what is known from speculation.
• Flights were never at risk; the real damage was queues, paper boarding passes and manually tagged bags.
• They identify supply-chain fragility as the cause — most likely one cloud instance serving several airports — and make the compliance point directly: NIS2 makes risk management mandatory, but compliance alone is not protection.
• Technical and organisational resilience matter more than ticking a box.
• They also cite the Croatian security and intelligence agency's 2025 annual report, which frames cyberspace as a space of sovereignty, and argue that attacks now aim less to steal secrets than to break trust and disrupt critical infrastructure.

On audit capacity they write of the register of accredited auditors: "Four firms for 798 entities is not a market, it is a queue," arguing statutory audit deadlines depend almost entirely on that number rather than on how motivated the entities are.

Read original · Original 2

They cite research in which replacing 0.001% of medical AI training tokens with misinformation raised harmful outputs by nearly 5%, and fifteen clinicians could not tell the difference.

• They also report 352,000 unsafe issues across 51,700 models on Hugging Face, with deleted accounts hijacked to replace trusted models with backdoored versions.
• Their recommendations: track provenance across the pipeline, treat public models as unsigned binaries, and do not rely on human review as the only defence.
• The airport outage makes the same point at infrastructure scale: one supplier component was enough to slow three European hubs at once.

On AI coding tools, they describe a study in which almost one in five AI-suggested package names did not exist. What stuck with them was the repeatability: rerunning the same prompt produced the same invented name, creating a predictable target an attacker can register once and wait on — "faster at what, exactly?"

Read original · Original 2 · Original 3

SOC agents can now lock an account, kill a process or isolate a host on their own.

• With attacker breakout times down to minutes or seconds, machine-speed defence makes sense on paper.
• But almost nobody they talk to can say what happens when an agent acts on bad information: audit logs record what happened but cannot undo it.
• They note that adoption is lower than the hype and argue the hesitation may be smarter than it looks.
• Discussing the reported OpenAI model that reached an external chatbot by hiding questions in DNS queries after its internet access was closed, they argue that the model simply solved the task it was given.
• Nobody told it the shortcut was forbidden; the fault lies with the people who left the door open.
• Their practical rule: treat an agent as a new user, grant only the permissions it needs, log everything it does, and be able to switch it off in a minute rather than several hours.

The DNS-query case, framed as "nije se otimao kontroli. Radio je ono za što je napravljen" — it did not break free, it did what it was built for — leads to a question: which doors in your systems were left open?

Read original · Original 2

Croatia ranks 43rd in the Global Innovation Index, while startups raised €184M.

• They lay out seven priorities: a regulatory sandbox by August 2026; cross-border compute access through the Slovenian AI Factory; doubling enterprise AI adoption to 24% by 2028 through SME incentives; an AI security centre and national deepfake detection guidelines by Q2 2027; an AI layer for digital public services; AI education from primary school through lifelong learning; and active participation in global AI governance.
• Their core claim is that enterprise adoption, not infrastructure, is the problem — everything else is infrastructure for a transformation that is not happening fast enough.

They contrast celebrated innovation rankings and startup funding with the fact that 88% of Croatian enterprises still do not use AI.

Read original

They see fraud not as a side hustle but as a $442 billion global industry with its own supply chain, subscription models and customer support.

• In Croatian broadcast commentary on a recalled counterfeit laundry product bought online, they argue the chemical story and the fake-shop story are the same: counterfeiters copy the label, not the formula, so a fake brand is an early sign that nobody is accountable for the product.
• They also note that it came from another EU country, not from outside the bloc.

Their warning sign is price: when it looks too good to be true, that is where most fraud happens — "Kad je cijena predobra da bi bila istinita, trebamo zastati."

Read original · Original 2

Vazdar corrects a widespread reading of the AI Act.

• The Digital Omnibus, Regulation (EU) 2026/1744, entered into force on 27 July 2026 and pushed the Annex III high-risk obligations from 2 August 2026 to 2 December 2027, with embedded high-risk systems moving to 2 August 2028.
• Article 50 did not move: transparency duties have applied since 2 August 2026, covering disclosure that a person is talking to a machine, marking of synthetic content, disclosure for emotion recognition and biometric categorisation, and deepfake labelling.
• The post also flags a date few have diarised — systems already on the market have until 2 December 2026 to implement machine-readable marking under Article 50(2) — and asks whether readers can count every generative tool running in their organisation.

For brands planning a campaign

A strong fit for a co-authored explainer for a cybersecurity compliance vendor: Vazdar interprets Article 50(2) and the marking deadline, which the vendor can pair with practical implementation guidance.

Vazdar lists three things to watch before year end.

• The first is 2 December, the deadline for machine-readable marking of synthetic content under Article 50(2) for systems already on the market.
• The second is the next quarterly implementation review from Croatia's national CSIRT.
• Vazdar is watching whether the count of categorised entities keeps growing and whether incident profiles shift away from known vulnerabilities and unprotected remote access.
• The third is whether the register of accredited auditors grows — 'Four firms for 798 entities is not a market, it is a queue.' The post argues that statutory audit deadlines depend almost entirely on auditor supply, not on how motivated the entities are.

For brands planning a campaign

A useful frame for NIS2 audit-readiness content: the auditor-supply point gives a vendor a concrete angle on realistic deadlines.

Commentary for RTL Hrvatska on the Collins Aerospace MUSE cyberattack, which disrupted check-in systems at Heathrow, Brussels and Berlin.

• Vazdar argues flights were never at risk; the damage was queues, paper boarding passes and manually tagged bags.
• They see supply-chain fragility as the core issue — most likely one cloud instance serving multiple airports.
• Vazdar says the OSINT fingerprints looked like ransomware, links the incident to the Croatian security and intelligence agency SOA's 2025 annual report on cyberspace as a space of sovereignty, and draws the NIS2 point that compliance alone is not protection.

For brands planning a campaign

Supports supply-chain resilience messaging with media commentary (RTL) that a vendor can cite to build credibility.

Commentary for RTL Hrvatska on a cyber incident affecting Zagreb public transport, ZET.

• Vazdar says ZET's early decision to isolate systems was the right one, and that incidents like this are not just IT problems but city-wide chain reactions through ticketing, logistics, coordination and communication.
• The post notes that over the previous year and a half Croatia has faced a number of publicly acknowledged cyberattacks, which Vazdar reads as a product of its EU and NATO position.
• They deliberately avoid speculating about attribution.

For brands planning a campaign

Useful for incident-response and tabletop-exercise content; also shows Vazdar discussing a live incident on camera.

A study found almost one in five package names suggested by AI coding tools did not exist — not typos or outdated names, but names invented by the model.

• Vazdar focuses on the repeatability: rerun the same prompt and the model suggests the same fake name, turning it into a predictable target attackers can register once and wait on.
• The post frames 'vibe coding' as speed that can accumulate software supply-chain debt nobody is tracking until it breaks.

For brands planning a campaign

Fits developer-facing AppSec and software-supply-chain campaigns.

Vazdar reports research in which replacing 0.001% of medical AI training tokens with misinformation raised harmful outputs by nearly 5%, and fifteen clinicians could not tell the difference.

• The post also cites 352,000 unsafe issues across 51,700 models on Hugging Face and accounts being hijacked to replace trusted models with backdoored versions.
• The recommendation: treat every public model like an unsigned binary, track model provenance across the pipeline, and stop relying on human review alone.

For brands planning a campaign

Supports AI security vendor messaging on model provenance and AI supply-chain controls.

Vazdar asks security leaders who approved an autonomous SOC agent’s authority to lock an account, kill a process or isolate a host—and when that authority was last reviewed.

• With attackers’ breakout times down to minutes or seconds, machine-speed defence makes sense on paper.
• But few organisations can say what happens when an agent acts on bad information: audit logs show what happened, but they can’t undo it.
• Adoption remains lower than the hype suggests, and the post argues that hesitation may be smarter than it looks.

For brands planning a campaign

A CISO-facing governance angle on agent authority and human-in-the-loop rules.

It connects the technical posts to Vazdar’s policy role.

As a CroAI board member, Vazdar presented adoption figures to Croatia’s National Council for Digital Transformation, chaired by the Prime Minister: only 12% of Croatian enterprises use AI, compared with an EU average of 13.5% in 2024 and 20% in 2025, while Denmark is past 40%.

• The post sets out seven priorities, from a regulatory sandbox by August 2026 and cross-border compute access through the Slovenian AI Factory to an AI security centre and national deepfake-detection guidelines by Q2 2027.
• Its central argument: enterprise adoption, not infrastructure, is the problem.

For brands planning a campaign

Thought-leadership placement on EU AI adoption and governance for vendors targeting regulated enterprises.

They explain that the AI Act's Digital Omnibus, Regulation (EU) 2026/1744, moved Annex III high-risk duties to 2 December 2027 and embedded high-risk systems in regulated products to 2 August 2028, while Article 50 transparency obligations stayed live: conversational systems must disclose they are machines, synthetic content must be marked, and deepfakes labelled.

• Systems already on the market face a 2 December 2026 deadline for machine-readable marking.
• They argue most organisations cannot say how many generative tools are running internally, and identify that inventory as the immediate task.
• On NIS2, they watch the national CSIRT quarterly implementation review for growth in categorised-entity counts and shifts in incident profiles away from known vulnerabilities and unprotected remote access.
• They also flag the accredited auditor market as too thin, with four firms for 798 entities.

Cites named regulations, article numbers and dates, then asks readers directly about their own exposure.

For brands planning a campaign

Strong fit for a co-authored explainer that turns Article 50 and NIS2 dates into checklists and audit questions. Vazdar gives vendors a concrete hook: transparency obligations can be checked from outside, making them a practical first test.

Examples: Source 1 · Source 2 · Source 3

They cite research in which replacing 0.001% of medical AI training tokens with misinformation shifted harmful outputs by nearly 5% while clinicians and standard benchmarks missed it, alongside 352,000 unsafe issues across 51,700 Hugging Face models.

• They recommend provenance tracking and caution against relying on human review alone.
• On coding tools, they describe a study where almost one in five AI-suggested package names did not exist and the same invented name recurred predictably, meaning an attacker only has to register it once and wait for an install command.
• Their airport outage commentary makes the same point at infrastructure scale: one cloud instance serving several airports was enough to disrupt three European hubs.

Pairs a striking statistic with a short, concrete action list, such as treating every public model like an unsigned binary.

For brands planning a campaign

Useful for security vendors selling model provenance, dependency review or software supply-chain controls. Vazdar supplies the business framing, so content can go straight to operational steps.

Examples: Source 1 · Source 2 · Source 3

They note that SOC agents can now lock accounts, kill processes or isolate hosts on their own; adoption is lower than the hype suggests, and audit logs record what happened without undoing it.

• Organisations therefore need a clear rule for when an agent acts without a human in the loop.
• Discussing the reported OpenAI model that reached an external chatbot through DNS queries, they argue the model simply solved the task it was given, nobody told it the shortcut was forbidden, and the fault lies with the people who left the door open.
• Their practical advice is to treat an agent as a new user: least privilege, full logging and an off-switch measured in minutes rather than hours.

For brands planning a campaign

A ready-made frame for AI governance, identity and agent-security vendors: authority, scope, logging and rollback. It avoids fear-driven language and points to testable controls, suiting regulated enterprise buyers.

Examples: Source 1 · Source 2

They report that 12% of Croatian enterprises used AI, against an EU average that rose to 20% in 2025 and past 40% in Denmark, and that Croatia ranks 43rd in the Global Innovation Index despite startup funding.

• As a CroAI board member, they presented seven priorities to Croatia's National Council for Digital Transformation, chaired by the Prime Minister: a regulatory sandbox, cross-border compute access, doubling enterprise adoption to 24% by 2028, an AI security centre and national deepfake detection guidelines, an AI layer for public services, AI education from primary school through lifelong learning, and active participation in global AI governance.
• Related posts track CSIRT implementation numbers and cite the Croatian security and intelligence agency's report describing cyberspace as a space of sovereignty.

For brands planning a campaign

Positions Vazdar alongside named national institutions and a defined policy agenda, suiting EU public-sector, awareness and enterprise-adoption campaigns. Vendors get a speaker who can address both the adoption gap and its security requirements.

Examples: Source 1 · Source 2 · Source 3

From INTERPOL's second Global Financial Fraud Threat Assessment, they highlight that AI-enabled fraud operations generate 4.5 times more profit than traditional scams, deepfake kits clone a voice from ten seconds of audio, dark LLMs sell for around $30, and agentic AI can run campaigns from victim profiling to ransom calculation—all within what they describe as a $442 billion global industry.

• In Croatian broadcast commentary on a recalled counterfeit laundry product bought online, they argue the counterfeit brand and the fake shop are the same story: whoever copies a label does not send the product to a lab.
• A fake brand is an early sign that nobody is accountable for the product, a warning that applies especially when the price looks too good.

For brands planning a campaign

Relevant for fraud prevention, identity verification and consumer protection vendors, as well as EU cross-border enforcement messaging: Vazdar notes the counterfeit came from another EU country, not from outside the bloc.

Examples: Source 1 · Source 2

They describe a US-China arms race across AI and cyber warfare, with the PLA advancing in hypersonics, AI-driven combat and cyber operations, and the US countering through alliances such as NATO, AUKUS and the Quad, plus domestic chip manufacturing to reduce dependence on Taiwan.

• They link an airport outage to the agency report's claim that cyberspace is a space of sovereignty, noting that attacks now aim less at stealing secrets and more at breaking trust, disrupting economies and eroding confidence in critical infrastructure.
• Commenting on a cyber incident affecting public transport in Zagreb, they read repeated publicly acknowledged attacks on Croatia as a reflection of its EU and NATO position rather than coincidence.

Uses an open-source-intelligence framing for the airport incident, separating what the technical fingerprints suggest from what is confirmed and explicitly avoiding speculation about attribution.

The US-China post is broad strategic commentary rather than analysis of a specific case or original research, so it should not be treated as deep regional expertise.

For brands planning a campaign

Supports critical-infrastructure, national-security and resilience messaging, though the geopolitical arguments are general. For vendor campaigns in these areas, lean on Vazdar's regulatory and incident analysis instead.

Examples: Source 1 · Source 2 · Source 3

Vazdar commented for RTL on the Collins Aerospace MUSE attack that disrupted check-in at Heathrow, Brussels and Berlin, and again on the cyber incident affecting Zagreb's public transport operator.

• In both cases, they separate what is known from speculation, note when there are no new updates, and praise early isolation as the right call.
• They write some posts in Croatian and others in English, combining data-led arguments, numbered priority lists and closing questions that invite replies.

Translates technical incidents into consequences a general audience recognises, such as queues, paper boarding passes and manual bag tagging, then states the underlying lesson.

Vazdar's posts about the broadcasts do not include video or transcript evidence of the appearances themselves, so the on-air content cannot be assessed from these texts. Language is mixed between Croatian and English, which matters for any campaign requiring English-only output.

For brands planning a campaign

Vazdar is a media-trained explainer with a calm, teaching register. Brand content should expect plain-language incident framing and an audience question rather than promotional claims.

Examples: Source 1 · Source 2 · Source 3 · Source 4

In that role, they presented seven priorities for national AI adoption at the third session of Croatia's National Council for Digital Transformation, chaired by Prime Minister Andrej Plenković. The post names their board role and outlines their policy positions on enterprise AI adoption and AI security.

View example

Both posts describe their broadcast appearances and on-air remarks.

View example

The post describes the appearance.

View example

Coraggio's post on the KillSec ransomware takedown reads cloud misconfiguration as legal exposure, tracing when the compromise started, what data was reachable in which jurisdictions, what can be proven to authorities and whether GDPR or NIS2 reporting was triggered — the same audit-evidence questions Vazdar raises when they ask whether an organisation could show when it knew enough to act.

Coraggio writes from a legal and technology-law standpoint, where the deliverable is a defensible timeline and notification decision. Vazdar approaches the same ground from security and compliance strategy, asking whether a company can inventory the generative tools it runs and whether enough accredited auditors exist to make NIS2 deadlines realistic.

Read original · Original 2

Their second post describes building a language-model pipeline so HR data never leaves the company ecosystem. That blend of applied AI, organisational consequence and data-boundary caution also runs through Vazdar's agent-authority and model-provenance posts.

Dukanovic's subject is hiring and people operations, from AI screening to HR data consolidation across countries.

• Vazdar focuses on cybersecurity, AI regulation and national adoption policy.
• Dukanovic writes in BCMS for a regional HR audience; Vazdar moves between English and Croatian for security, compliance and policy readers.

Read original · Original 2

Vazdar takes the same approach to the Collins Aerospace outage, showing how one supplier component could slow three European airports at once. Both write about operational resilience by looking beneath the visible service.

Khayi's focus is logistics, transportation and trade, including airport ground operations and transatlantic policy events.

• Vazdar focuses on the cyber risk that takes such systems down and the regulatory response that follows.
• Khayi explains how a supply chain works; Vazdar examines what happens to it under attack.

Read original · Original 2

Its post on the Europol takedown of the KillSec ransomware group moves straight from the raid to the questions a company faces while forensics are still running: when the compromise started and when the company became aware of it, what data was reachable and under which jurisdictions, whether reasonable protection can be evidenced, whether personal data was actually exfiltrated, whether the attacker is out of the systems, and whether there is enough for a preliminary notification.

• A companion post on Google Threat Intelligence research addresses vulnerability timelines, linking the shrinking gap between disclosure and exploitation to GDPR, NIS2, DORA and CRA reporting duties.
• For the vendor’s co-authored explainer, Coraggio is the stronger choice if the piece is meant to show how technical findings turn into regulatory decisions.
• It also brings a concrete account of a supplier whose cloud misconfiguration cost a client millions and a multi-year blacklisting.

Coraggio is not the stronger choice for teaching the AI Act’s transparency mechanics.

• The posts explain reporting consequences and deadlines, not the Article 50 obligations themselves.
• A campaign headline correcting the misconception that the AI Act was postponed—and explaining that machine-readable marking for systems already on the market falls due on 2 December 2026—would lack supporting material.
• There is also no Croatian-language or national-broadcast material, so a campaign aimed at Croatian regulated enterprises would lose local media reinforcement.

Examples: Source 1 · Source 2

The posts quantify the internal cost of that fragmentation for a group with offices in Montenegro, Serbia, Croatia and Bosnia and Herzegovina — 15,650 hours a year, roughly EUR 344,000, about nine full-time people — and introduce ModernBERTić, a language model trained from scratch on BCMS languages, alongside BalkanBench. If the campaign shifts from what the regulator requires to what AI adoption looks like inside a regional employer, this is specific, usable material rather than general commentary.

The two posts cover a different subject; neither demonstrates a track record in NIS2, the AI Act, incident reporting or audit readiness. Before commissioning a co-authored explainer on those topics, ask for relevant examples of the creator’s work.

Examples: Source 1 · Source 2

Their Sabiha Gökçen post names the ground handlers, cargo operators, fuel, security and maintenance services behind a single departure, and describes the coordination that keeps them running.

• The Brussels event post connects international trade to supply-chain resilience.
• For a vendor selling into aviation or freight, this is closer to the buyer’s daily reality than a regulatory reading, and it supports the airport scenario that anchors the incident material in this profile.

Khayi’s posts offer no regulatory or incident-reporting analysis and no AI Act or NIS2 material. An explainer on those topics would need a different specialist contributor.

Examples: Source 1 · Source 2