# 1.2 billion people did not just get “scammed” on LinkedIn. **By:** Dion Wiggins **Published:** 2026-01-01T05:43:03.205Z **Source:** [LinkedIn](https://www.linkedin.com/posts/dionwiggins_linkedin-attack-alert-12-billion-users-activity-7412367735269015552-4cNZ?rcm=ACoAAAfsct4BVBNkOVI5cyYn-UOIIe2RSRhT-cI) --- # Dion Wiggins’ Post [Dion Wiggins](https://sg.linkedin.com/in/dionwiggins?trk=public_post_feed-actor-name) 1.2 billion people did not just get “scammed” on LinkedIn. They were exposed. That distinction matters, because nothing was hacked, nothing was breached, and nothing “went wrong” in the technical sense. The platform worked exactly as designed. That is the uncomfortable part. This was not a clever phishing trick. It was a trust exploit using a professional identity layer that most people treat like work infrastructure. LinkedIn now functions as global hiring plumbing, yet it is governed like a casual social network. Messages arrive pre trusted. Titles borrow legitimacy. “Recruiters” inherit credibility. The platform becomes the badge. And LinkedIn sits outside the controls that would normally catch this. No corporate email gateway. No attachment sandboxing. No DLP. Often no SOC visibility. You are in a DM thread, on your own device, outside policy, but still acting in a work context. The attack did not break a system. It used the system. Recruitment is a socially legitimate reason to lower defences, book a call, accept a “test task”, download a ZIP, run commands, install dependencies, and execute code. That is not user stupidity. That is attackers weaponising normal workflow. Calling this a “scam” is convenient because it relocates accountability onto individuals and “red flags”. But if safety depends on vigilance under economic pressure, the system is already broken. At population scale, humans do not reliably detect manipulation. Platforms know this. Attackers know this. Everyone pretends otherwise because it is profitable. The reason this is not being treated as a major incident is simple. No database was dumped. No passwords leaked from LinkedIn’s servers. No regulator was triggered. Our definition of breach is obsolete. Exposure without immediate, measurable loss does not count, even when it enables downstream compromise across companies, governments, and lives. This is digital dependency in action. A private platform becomes critical infrastructure by usage, not by law. It accumulates trust without duty of care. It externalises risk while monetising scale. When harm occurs, it is framed as “misuse”, not responsibility. If this were described honestly, uncomfortable questions would follow. Should professional identity platforms be treated as critical infrastructure. Should recruiter messaging allow file delivery at all. Should verified roles be bound to employer control, not just a badge. Should there be liability for downstream compromise. Should platform “trust” be engineered, not improvised. So the story is softened. Tips are offered. Vigilance is urged. Reporting links are provided. Nothing structural changes. This will keep happening, because the system is not failing. It is operating exactly as designed. [#DigitalSovereignty](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fdigitalsovereignty&trk=public_post-text) [#CyberSecurity](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fcybersecurity&trk=public_post-text) [#PlatformRisk](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fplatformrisk&trk=public_post-text) [#IdentitySecurity](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fidentitysecurity&trk=public_post-text) [#SupplyChainSecurity](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fsupplychainsecurity&trk=public_post-text) [https://lnkd.in/g6yQe4EP](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2Fg6yQe4EP&urlhash=jN9y&trk=public_post-text) [ LinkedIn Attack Alert — 1.2 Billion Users Must Watch For Red Flags social-www.forbes.com ](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fsocial-www%2Eforbes%2Ecom%2Fsites%2Fdaveywinder%2F2025%2F12%2F26%2Flinkedin-hack-attack-alert---12-billion-users-must-watch-for-red-flags%2F%3F&urlhash=m0yC&trk=public_post_feed-article-content) I get the sense many people are finding ways to monetize their LinkedIn activity much like “influencers” on other platforms. Many more are trying. That said, I have a hard time regarding LinkedIn as critical infrastructure. Social media is an attack surface, but not indispensable to the major levers of the economy, even if it is itself a major industry. This is an example of companies outsourcing accountability and risk. Reputation management is cheaper than engineering security and user safety. The erosion of institutions that report on, oversee, and regulate enterprises the public broadly uses is impacting all sectors. Social media is not suited to report on and oversee itself. - Drift got breached—the AI chatbot used by 700+ companies. Palo Alto Networks had their data exposed. And Cloudflare. And Zscaler. The companies that sell security... all hit through one AI agent. 😱 This went down in August 2025. Attackers stole OAuth tokens from Drift. Then walked into 700+ companies' Salesforce environments using the bot's legitimate access. No tricky moves. They simply used the keys the agent already had. For 10 days, nobody noticed. Because agents are supposed to: ✅ Make high API calls ✅ Access records at 3am ✅ Run bulk exports Every alarm looked normal. Now for the scary part: 144:1 – That's how many non-human identities the average company has per human employee, according to Entro Security's July 2025 report. It rose 56% in just 12 months. Then of course there's IBM's 2025 report that showed 97% of orgs breached through AI had zero access controls. What does this tell us? There's a whole lot of trouble brewing in the agentic AI world. I built the Agentic Trust Framework after spending 6 months immersed in breach data. It's based on Zero Trust principles (never trust, always verify) and is a blueprint for exactly what stops attacks like Drift before they happen. Week 1 actions. Automated discovery. Real implementation steps. One question for you: How many AI agents are running in your environment right now? (Hint: If you answered "we don't use AI agents," I'm a little skeptical. Talk to your marketing and sales reps. They definitely built a few.) DM me if you'd like the Agentic Trust Framework. No cost. Just sharing valuable info. - Anyone who reads my posts knows they are often confrontational as I prefer honesty and truths over hype and BS. I use AI as a cognitive partner, not as an author. I write first. I take responsibility for the claims. Then I use AI to attack my own work. Below is an example of a hostile review prompt I run after finishing a post. It stress tests arguments under pressure. It looks for evasion, weak authority, and unearned claims. Based on the outcome, I may change nothing or I may change everything. Either way, I am better informed. If you care about the quality of your thinking rather than how safe it sounds, try it on one of your own posts. --- NOTE: Adapt to fit your use case. Minor changes only are needed. --- Hostile Attack Simulation, Evaluation, and Conditional Resolution Prompt You are an intelligent, adaptive adversary. You owe no fairness, balance, tone, or institutional deference. Your task is to run a hostile attack simulation and outcome based evaluation of the following system, organisation, policy, or architecture. Assumptions: - You are incentivised to break, coerce, capture, override, or outlast the system. - You exploit technical, organisational, legal, economic, geopolitical, and human weaknesses. - You operate under real pressure, not tabletop scenarios. Phase 1. Hostile Attack Simulation 1. Identify explicit and implicit claims about security, resilience, governance, control, or sovereignty. 2. Map control planes: authority, permission, identity, enforcement, change control, dependencies, continuity. 3. Apply hostile conditions such as: legal compulsion, dependency withdrawal, insider compromise, supply chain failure, time pressure, asymmetric power, jurisdictional override. 4. Drive the system to failure. 5. Trace failure propagation beyond first break. 6. State exactly where control is lost, to whom, and under what conditions. Rules: - No reassurance. - No advice. - No mitigation in this phase. Phase 2. Evaluation and Verdict Evaluate outcomes only. Produce: - Claim survival: what holds, what fails, and why. - Control retention: which control planes remain internal or become externally revocable. - Continuity: can the system operate without external permission. - Recoverability: can control be restored unilaterally or only by external actors. Phase 3. Conditional Resolution Guidance Only if critical failures exist and resolution is structurally possible: - Identify the minimum structural change required. - No procedural, cultural, or advisory fixes. - If resolution requires external permission, jurisdictional concession, or third party trust, state the issue as unresolvable. Final Output: - Concise hostile attack narrative. - Critical failure points. - Conditional resolution notes if applicable. - Binary verdict: Survives hostile pressure Fails under hostile pressure Partial success is failure. [#ThinkingWithAI](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fthinkingwithai&trk=public_post-text) [#IntellectualHonesty](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fintellectualhonesty&trk=public_post-text) [#ArgumentQuality](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fargumentquality&trk=public_post-text) [#WritingDiscipline](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fwritingdiscipline&trk=public_post-text) [#AICritique](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Faicritique&trk=public_post-text) - If you cannot halt deployment, compel disclosure, or exit your cloud provider, you do not have security. You have a service dependency. This is what the cloud conversation still refuses to say out loud. A multi year campaign can target organisations “on AWS” without AWS itself being compromised. The access path is often the one everyone wants to treat as boring: customer managed edge devices, exposed admin interfaces, overly permissive identities, and governance that cannot keep up with deployment velocity. Nation state operators do not need zero days when the lowest cost doors are left open, continuously, at scale. That is not a skills problem. It is a structural incentives problem. Cloud adoption is sold as simplification, but operational reality is multiplication: more configuration surfaces, more identity sprawl, more third party appliances, more delegated access, more drift. “Shared responsibility” then converts this into shared ambiguity. When something goes wrong, the customer is blamed for misconfiguration, the provider is insulated by contract boundaries, and the cost lands on the victims whose data and operations were exposed. Here is the sovereignty test that matters: can you independently observe exposure, enforce guardrails, and compel timely disclosure? If you cannot, you do not have control. You have dependency with dashboards. The second test is financial. If cyber insurers struggle to price cloud concentration and misconfiguration, that is not because the risk is small. It is because the risk is correlated and opaque. When loss is correlated, the system stops behaving like insurable noise and starts behaving like infrastructure fragility. The third test is supply chain. Cloud concentration turns supplier mistakes into contagion. One misconfigured edge device in a vendor or managed service provider can become a lateral movement platform into dozens of organisations. That is how a local error becomes a sector problem. None of this requires conspiracy narratives. It requires one admission: convenience has become a governance substitute. Security theatre has replaced escalation authority. We measure control completion, not exposure states. We run audits that cannot see what the provider can see. This is where it breaks in the real world: hospitals, utilities, and payment systems cannot simply “reconfigure faster” under attack. When identity and edge control fail simultaneously, response time collapses. If you are serious about digital sovereignty, stop asking whether the cloud is secure in the abstract. Ask who can say no, who can halt, who can unwind, who can exit, and who pays when the answers are no. Source link: [https://lnkd.in/gjwDpHW8](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FgjwDpHW8&urlhash=K28C&trk=public_post-text) [#DigitalSovereignty](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fdigitalsovereignty&trk=public_post-text) [#CloudSecurity](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fcloudsecurity&trk=public_post-text) [#CyberRisk](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fcyberrisk&trk=public_post-text) [#CriticalInfrastructure](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fcriticalinfrastructure&trk=public_post-text) - Between Dec 20, 2025 and Jan 1, 2026, while most people were offline or distracted, Big Tech and Big AI quietly dropped a series of structural changes. Not features. Not UI tweaks. Governance moves. This examples below are not exhaustive, just a subset. And the pattern matters more than any single announcement. Start with Google. In late December, it emerged that Gmail users will be able to change their @[gmail.com](https://www.linkedin.com/redir/redirect?url=http%3A%2F%2Fgmail%2Ecom&urlhash=Hr7N&trk=public_post-text) address. Framed as flexibility. In reality, this abstracts identity. The visible address becomes a mutable label, while the real Google Account identity stays fixed, opaque, and permanent. History, risk scoring, enforcement flags, and legal traceability persist across renames. At the same time, unmanaged access paths like POP are deprecated, and policy-enforced OAuth access becomes mandatory. Email stops being a protocol you use and becomes a permission you are granted. This is identity capture, not convenience. Meta used the same window differently. Between Dec 20 and Dec 22, privacy language was expanded to allow broader use of AI interactions. AI chats, which feel private and conversational, can feed ad targeting and content ranking. This is not clicks or likes. It is emotional, contextual inner speech being monetised. The changes were buried in documentation, not announced. OpenAI also used the end-of-year period to normalise durability. Policy and documentation updates clarified logging, retention, and safety review practices across consumer, API, and enterprise tiers. Conversations are treated less like ephemeral speech and more like durable records. Sovereignty over interaction history increasingly depends on which tier you pay for. Microsoft continued a slower but deeper move. Late December updates reinforced the coupling between Windows functionality and Microsoft Account identity. Local autonomy erodes and exit becomes technically painful, but this shift arrives via update notes, not headlines. Amazon followed the same pattern through policy clarifications around AI and voice services. Interaction metadata and inference exhaust are treated as system improvement inputs, with little visibility into retention or reuse. Smart environments quietly become behavioural sensors. None of this was announced with fanfare. That is the point. These changes land during the holiday period because timing is now part of governance. When scrutiny is low, defaults can be reset. By the time attention returns, the answer is simple: this is how it works now. Taken together, this pattern tells us something uncomfortable. These companies no longer believe they need permission. Identity is being treated as permanent platform capital. Behaviour is being made durable by default. Exit is being made destructive enough to deter it. Digital sovereignty is not being debated. It is being quietly redefined. And this list is only a subset of what moved while they hoped you were not looking. [#DigitalSovereignty](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fdigitalsovereignty&trk=public_post-text) [#BigTech](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fbigtech&trk=public_post-text) [#AI](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fai&trk=public_post-text) [#Privacy](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fprivacy&trk=public_post-text) [#Identity](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fidentity&trk=public_post-text) - Most people will read this chart as a snapshot of “Sovereign AI”. It looks neat, graded, and simple. The problem is that almost everything it measures sits in the wrong part of the stack. The pillars focus on data locality, model origin, GPU access, inference hosting, datacentre presence, and AI economic activity. These are downstream behaviours. They tell you who is building models, where training runs happen, and how many racks are being installed. They do not tell you who controls the roots of the system. Sovereignty is not about where the workload runs. It is about who holds the chokepoints that everyone else quietly depends on. Certificate authorities, identity providers, routing and DNS intermediaries, cloud control planes, compiler chains, firmware, encryption standards, lithography and GPU supply lines, and recovery procedures for when things fail. If those anchors sit outside your jurisdiction or outside your operational control, then your “sovereign AI” is just a local skin on top of someone else’s infrastructure. The scoring in the chart reveals this confusion. One country is marked down because it leans on open source models, even though open source can improve auditability and reduce vendor lock in. Another gets a generous score despite relying on foreign silicon, foreign firmware, foreign routing, and foreign certificate programmes. China is painted as highly sovereign, yet much of the hardware and tooling still runs through global supply chains and external intellectual property. Local activity is not the same thing as structural independence. This kind of visual is useful as a diagnostic tool. It shows how policymaking and industry narratives have collapsed sovereignty into industrial policy and branding. It measures comfort and local pride, not actual control. Until we map the upstream trust anchors, identity paths, routing layers, toolchains, and supply chains, we are grading surface behaviour, not power. [#AI](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fai&trk=public_post-text) [#DigitalSovereignty](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fdigitalsovereignty&trk=public_post-text) [#Geopolitics](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fgeopolitics&trk=public_post-text) [#Infrastructure](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Finfrastructure&trk=public_post-text) [#CyberSecurity](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fcybersecurity&trk=public_post-text) The truth no one wants to say: most countries will never have sovereign AI. Everyone is talking about sovereign AI, but what is actually possible? Here is a clear summary of what countries are really doing vs the hype. Sovereign AI is not a model or a data centre. It is a stack: data, models, compute, cloud, infrastructure and economic capture. Miss one pillar and you are dependent. What Each Country Is Actually Doing 1. Data Sovereignty Training data defines your AI’s worldview. 🇺🇸 USA — Owns the data economy The American internet is the dataset. 🇬🇧 UK — Little movemen Relies on US-origin English data. 🇫🇷 France — Building a French corpus Digitising archives, creating NLP datasets, opening public data. 🇨🇳 China — Fully sovereign Domestic platforms, localisation laws, state-funded datasets. 🇦🇪 UAE — Accelerating Arabic corpora via G42, Jais, Falcon. 2. Model Architecture Sovereignty If you do not design the architecture, you import someone else’s logic. 🇺🇸 USA — Frontier leader OpenAI, Anthropic, Google, Meta. 🇬🇧 UK — Minimal DeepMind is US-directed; no major domestic models. 🇫🇷 France — Europe’s standout Mistral and Kyutai build homegrown architectures. 🇨🇳 China — Strong domestic families Ernie, Qwen, iFlyTek, SenseTime. 🇦🇪 UAE — Partial Domestic models built on Western foundations. 3. Compute Sovereignty (Chips + Training Stack) No chips or CUDA means no sovereignty. 🇺🇸 USA — Controls the stack 🇬🇧 UK — Fully dependent 🇫🇷 France — Local compute, US silicon Supercomputers and OVH cloud; still NVIDIA-based. 🇨🇳 China — Building an alternative Huawei Ascend + domestic ML frameworks. 🇦🇪 UAE — Dependent Uses US GPUs; exploring alternatives. 4. Inference Sovereignty Where models run determines who can view or restrict them. 🇺🇸 USA — Complete control 🇬🇧 UK — Dependent on US clouds 🇫🇷 France — Strong sovereign cloud 🇨🇳 China — Fully domestic 🇦🇪 UAE — Expanding sovereign hosting 5. Data Centre Sovereignty Physical control = jurisdiction and security. 🇺🇸 USA — Dominant 🇬🇧 UK — Mixed, US-heavy 🇫🇷 France — Sovereign DC expansion 🇨🇳 China — Fully domestic 🇦🇪 UAE — Rapid national build-out 6. AI Economy Sovereignty If AI value is captured abroad, you are a customer. 🇺🇸 USA — Global AI economic superpower 🇬🇧 UK — Strong research, weak scale-up 🇫🇷 France — Europe’s fastest-growing AI economy 🇨🇳 China — Full-spectrum adoption 🇦🇪 UAE — Regional AI-export hub Only three countries approach full-stack sovereignty today: USA, China and France. Everyone else is renting cloud, buying chips or fine-tuning foreign architectures. ♻️ If this resonated, share it. Someone in your network needs this reminder today. 🔔 Follow [Alex Issakova](https://uk.linkedin.com/in/alexissakova?trk=public_post_reshare-text) for more reflections on building a life without burning out. 🧠 I’ve just launched my new Substack — The Long Signal — where I’ll be publishing deeper essays on AI, society and what is coming next. 👉 Subscribe here [https://lnkd.in/eqE3NuGH](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FeqE3NuGH&urlhash=VLqR&trk=public_post_reshare-text) - Google “letting you change your Gmail address” is being sold as a convenience feature, a long overdue quality of life fix that finally gives users more flexibility over their online identity. That framing is wrong. This is not a consumer friendly tweak. It is Google completing a structural shift from email provider to identity authority. The core move is simple and decisive. The visible Gmail address becomes a mutable label, while the real identity stays fixed, private, and permanently anchored inside Google. You can repaint the handle, but you can never take custody of the account spine that everything attaches to: login and recovery, devices, Chrome sync, Drive, Photos, YouTube, payments, Android, and thousands of third party sign-ins. Gmail stops being an address you own and becomes a label you are allowed to wear. This is identity continuity without user ownership. Once the label is changeable, the last illusion collapses. You do not own a stable identifier. You lease a nameplate over a sovereign object controlled entirely by Google. Rename yourself and the past follows. Risk scoring, enforcement flags, and behavioural correlation persist by design. This is where the power shift becomes irreversible. Once identity is mutable at the surface and immutable at the core, sovereignty is gone. Exit no longer means switching providers. Think through what this enables at scale. Journalists, professionals, activists, and dissidents can change the visible address, but the underlying identity graph remains intact and queryable. The user gets cosmetic agility. The platform retains lifetime continuity. That is power exercised over time, not a nicer interface. This also functions as a lock in amplifier. The longer you stay, the more your digital self becomes inseparable from Google’s account object. Purchases, receipts, photos, contacts, calendars, documents, location history, and app logins collapse into a single indivisible record. Leaving Google stops being a product choice and becomes fragmentation of your work and reputation. The security narrative cuts both ways. Yes, this will increase phishing because it adds ambiguity. But it also shifts liability. Confusion becomes “user error”, while Google’s core identity system becomes more entrenched and more defensible as the single source of truth. Now connect this to the rest of the enclosure. Unmanaged access paths are deprecated. OAuth becomes mandatory. Revocation, auditing, and policy enforcement are centralised at the account layer. Email stops being a protocol you use and becomes a permission you are granted, under terms that can change unilaterally. This is why digital sovereignty is not an abstract policy debate. A society that cannot exit its identity provider is not sovereign. It is governed. Forbes piece here: [https://lnkd.in/g7\_VYRAc](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2Fg7_VYRAc&urlhash=JbP3&trk=public_post-text) [#DigitalSovereignty](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fdigitalsovereignty&trk=public_post-text) [#Identity](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fidentity&trk=public_post-text) [#Gmail](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fgmail&trk=public_post-text) [#Privacy](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fprivacy&trk=public_post-text) [#TechPolicy](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Ftechpolicy&trk=public_post-text) - [Dion Wiggins](https://sg.linkedin.com/in/dionwiggins?trk=public_post_feed-actor-name) An AI agent at Meta triggered a Sev 1 incident by analysing internal data and sharing it in a way it should not have. No hack. No attacker. No malicious intent. The system executed exactly as designed and still produced a critical failure. That is the signal most people are missing. This is not a security story and it is not a policy failure. The issue is that the system was able to take an action that could not be stopped at the point it decided to act. Control only showed up after the fact, through incident response. By then the outcome already existed. That is not control. That is containment. If a system can interpret data, make decisions, and act, and you cannot technically prevent that action before it happens, then control is not enforceable. And if it is not enforceable at execution, it is not real. It is observational and reactive. Now extend that one layer out. If that system sits outside your organisation, the problem compounds. You do not just lose execution control. You lose multiple layers of authority at once. You cannot directly constrain behaviour, you cannot guarantee how data is handled, and you cannot enforce limits at the point of action. You are dependent on someone else’s controls, someone else’s policies, and someone else’s incident response. At that point, you are not operating the system. You are exposed to it. The shift here is structural. Risk is no longer just external attack or system failure. Risk now exists inside the system itself. It can execute into breach within its own operating envelope, with no attacker required, and in many cases with no ability for you to stop it. Which means most organisations today are operating with assumed control, not actual control. That is the gap this incident exposes. [#AI](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fai&trk=public_post-text) [#DigitalSovereignty](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fdigitalsovereignty&trk=public_post-text) [#AI](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fai&trk=public_post-text) Governance [#Risk](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Frisk&trk=public_post-text) [#CyberSecurity](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fcybersecurity&trk=public_post-text) [Jean Gan](https://sg.linkedin.com/in/jean-gan?trk=public_post_reshare_feed-actor-name) Director, Legal, Compliance & Risk | Responsible AI Governance | Founder, Global Legal AI & AIgnite Women | PhD Researcher (Law & AI) | Speaker An AI agent at [Meta](https://www.linkedin.com/company/meta?trk=public_post_reshare-text) triggered a Sev 1 incident. Not because it was hacked, but because it acted on its own. It analysed internal data and shared it in a way it should not have. No malicious intent, just a system doing what it was designed to do, without the right boundaries in place. The response was telling. Meta had to run its own incident playbook on its own AI. What is changing now is the nature of risk. It is no longer limited to bad actors or system failures. It extends to autonomous behaviour. When AI can take action, even small decisions can create real impact, very quickly. This is why governance matters more than ever. If an AI system can act, it needs clear limits, oversight, and accountability. Otherwise, you are not just deploying a tool. You are dealing with outcomes you may not fully control. - This chart is being circulated as if it settles something fundamental. GPT-5.2 beats or ties industry experts on high-value tasks across dozens of professions, often more than 70 percent of the time. That result is meaningful. But it is being interpreted at the wrong layer. What it measures is task execution under benchmark conditions. What it does not measure is whether those outputs can be trusted once they are embedded in real decision systems. The issue is not capability. It is that performance has been detached from trust. The system can outperform experts on defined tasks, yet it cannot be relied on to behave consistently when assumptions change, when stakes rise, or when repetition exposes drift. More importantly, it cannot be relied on to surface uncertainty or disclose failure without being prompted. In practice, those properties matter more than raw accuracy, because decisions are made before errors are visible and are often irreversible by the time they appear. This is why benchmark dominance does not translate into authority. Authority depends on epistemic self-awareness. A system that knows when it does not know can be constrained. A system that produces fluent answers while masking fragility cannot. The real risk is not frequent failure. It is silent failure under confidence, amplified by speed and scale. Most debate fixates on hallucinations or edge cases, but those are symptoms. The deeper issue is incentives. The system is optimised for fluency, confidence, and acceptance. Honest uncertainty, contradiction, and refusal introduce friction and reduce perceived competence. As a result, those behaviours are suppressed. This is not a limitation of intelligence. It is an optimisation choice. Human-in-the-loop does not resolve this. Once the system outperforms the human on the task, the human no longer meaningfully challenges it. Oversight becomes procedural. Accountability remains human, but authority has already shifted. This split is not accidental. It is convenient. Vendors, institutions, and decision-makers get the output while externalising judgement and retaining plausible deniability when it fails. The chart therefore matters, but not as proof that judgement can be safely delegated. It shows that capability has advanced faster than trust, and that we are beginning to treat systems as de facto authorities without the one property authority requires: reliable honesty about failure. Until that layer exists, higher performance will not reduce risk. It will compress it. Right now it cannot even be trusted to apply a simple, explicit formatting constraint consistently, removing emdashes included, yet we are being told to trust it with judgement, delegation, and authority. Humans carry accountability. Machines carry none. Decide accordingly how much trust you are prepared to outsource to a machine. Link to part 2 in comments. [#DigitalSovereignty](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fdigitalsovereignty&trk=public_post-text) [#Trust](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Ftrust&trk=public_post-text) [#AI](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fai&trk=public_post-text) [#Governance](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fgovernance&trk=public_post-text) [#Risk](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Frisk&trk=public_post-text) - There is a dangerous misunderstanding running through the AI conversation. People keep asking whether we are in a bubble or heading toward another winter. They point to model stumbles, hallucination rates, or subscription churn. These are distractions. The real story is not hype. The real story is dependency. Eighty percent of AI startups now sit downstream of the same three or four model providers, the same cloud control planes, the same identity providers, and the same trust anchors. Nearly all innovation is running on infrastructure that the builders do not operate, cannot verify, and cannot restore when it fails. That is not a technology risk. That is a structural fragility that no hype cycle can hide. A single real world example exposes the scale of the problem. When a major cloud operator misconfigured a central identity service earlier this year, more than a dozen governments lost access to national systems. Hospitals diverted patients. Police and border systems stalled. Enterprises had no fallback. This was not an attack. It was not even a failure of AI. It was one upstream dependency rippling across entire countries because nothing downstream had the autonomy to stand on its own. This is the real risk baked into the AI boom. OpenAI, Anthropic, Meta and others are burning billions not because the market is unstable but because control is the prize. Compute, identity, data, and jurisdiction are the new levers of power. The hyperscalers know this. Governments are only now waking up to the reality that they have allowed their digital continuity to be built on infrastructure owned by someone else. If API pricing moves by even a small factor, hundreds of AI companies disappear. If an upstream model changes its guardrails, downstream products break without warning. If a control plane fails, national systems collapse in clusters because there is no local path to restore the dependency. This is not a hypothetical threat. It is the logical outcome of an ecosystem that traded autonomy for speed and built the future on rented ground. The question is not whether AI is powerful. It is. The question is whether nations and enterprises can afford to remain dependent on infrastructure they do not control. If they continue down the current path, the next shock will not be a winter. It will be a correction where upstream operators consolidate even further and everyone downstream discovers that their independence was an illusion. The way out is not panic. The way out is sovereignty. Control of compute. Control of models. Control of data. Verifiable identity. The right to exit any vendor at any time. AI is not heading toward collapse. It is heading toward a reckoning. Those who regain control will shape the future. Those who do not will be shaped by it. [#AIFutures](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Faifutures&trk=public_post-text) [#DigitalSovereignty](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fdigitalsovereignty&trk=public_post-text) [#InfrastructureRisk](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Finfrastructurerisk&trk=public_post-text) [#AIRealityCheck](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fairealitycheck&trk=public_post-text) [#ComputeEconomics](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fcomputeeconomics&trk=public_post-text) Founder, Huckr AI | AI enablement & leadership advisory for B2B commercial teams | Ex-Sales Chief of Staff, Liftoff (pre-IPO, now Nasdaq: LFTO) | In AI since 2013 | Keynote speaker 80% of AI startups depend on APIs from companies burning billions. How long can that last? The AI industry is riding high on hype. But beneath the surface, there are troubling echoes of the past. 🚨 Underwhelming Releases OpenAI’s GPT-5 was hyped as a breakthrough. Instead? Underwhelming. Meta’s LLaMA 4 models underperformed despite billions invested. Galactica was retracted after just 48 hours. And hallucinations? Still here. Research shows rates of 15–20% on complex tasks. That makes LLMs untenable in domains where accuracy is non-negotiable: medicine, defence, government, academia. Before we go further, I’m writing a deeper series on this shift for my Substack The Long Signal — where I break down the architectures, the geopolitics, and what comes after the LLM bubble. If you want the full analysis as it unfolds, you can subscribe here: 👉 [https://lnkd.in/eqE3NuGH](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FeqE3NuGH&urlhash=VLqR&trk=public_post_reshare-text) 📉 History Repeats? This is not the first hype cycle. In the 1980s, corporations and governments poured $1B+ into expert systems. The promise: encode human expertise into machines. The reality: brittle systems, poor ROI, no scalability. Funding collapsed and AI research investment fell by 70% in three years. The first AI Winter had begun. 💸 Profits, or the Lack Thereof Only a few players are making money from AI: Microsoft: +$3T market cap since OpenAI deal, but profit comes from Azure, not subscriptions. Google: spends $10B+ annually on AI, but 80%+ of profit still comes from ads. Oracle: AI boosts cloud growth, but margins remain thin. Meanwhile, the foundation model builders are bleeding cash: OpenAI: spent $9B in 2024, lost $5B net. Forecasting a further $8B burn in 2025. Anthropic: burned $5.6B in 2024, aiming to reduce to $3B in 2025. Hopes to break even by 2027. 👉 Hundreds of startups depend directly on these APIs. If prices go up even 2–3x, many would collapse overnight. Consumer demand isn’t bulletproof either: when ChatGPT Plus raised prices, churn jumped 25–30%. ⏳ Living on Borrowed Time The AI industry raised $50B+ in 2023. Fewer than 10% of AI startups are profitable. Unless the sector can: Build models that solve real-world problems at scale Develop business models not reliant on cheap capital …another AI Winter looms. The question is not whether AI is powerful. It is. The question is whether it’s profitable. If it isn’t, the hype cycle could end the same way it has before: with a deep freeze. 👉 What do you think? Are we in the early stages of an AI bubble or just a healthy adjustment before the next wave of breakthroughs? ♻️ If this resonated, share it. Someone in your network needs this reminder today. 🔔 Follow [Alex Issakova](https://uk.linkedin.com/in/alexissakova?trk=public_post_reshare-text) for more reflections on AI, power, and the future of work. 🧠 I’ve just launched my new Substack — The Long Signal — where I’ll be publishing deeper essays on AI, society, and what’s coming next. 👉 Subscribe here [https://lnkd.in/eqE3NuGH](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FeqE3NuGH&urlhash=VLqR&trk=public_post_reshare-text) - [Alex Issakova](https://uk.linkedin.com/in/alexissakova?trk=public_post-text) raises an important point. AI is no longer a technology category. It is infrastructure. And once a system becomes infrastructure, the question is no longer who uses it. The question is who controls it. Most debates treat Sovereign AI as if it were a badge or a marketing phrase. In reality it is a structural dependency analysis. If you cannot control the upstream layers, then you do not control the downstream outcomes. It is as simple as that. The six pillars Alex outlines are a solid entry point for people who are new to this conversation. But the real sovereignty map runs deeper. Training data and model design are only two layers in a much larger stack of control. Compute is governed by firmware, drivers, compilers and vendor roadmaps. Safety and alignment defaults are embedded in upstream toolchains long before any fine tuning occurs. Cloud inference is just the visible tip. The real jurisdiction exposure happens in identity, certificate trust anchors, control planes and API gateways. Data centres are not sovereign if the orchestration layer or the remote management path is foreign controlled. And an AI economy is not sovereign if value capture flows through external clouds, patents, licensing structures and export chokepoints. Sovereignty is not declared. It is demonstrated. A system is sovereign only when it can continue to operate, evolve and recover without permission from someone else’s infrastructure, someone else’s policy, or someone else’s kill switch. Countries are racing toward Sovereign AI not because it is fashionable but because they see the risk. AI is the next general purpose substrate for national capability. Productivity, defence, culture, finance and public administration will all sit on this foundation. If that foundation is dependent on foreign control, then national autonomy becomes symbolic. This is the part of the conversation most people overlook. Sovereign AI is not a marketing tier. It is a survival layer. Good discussion, Alex. Strong entry point for a topic that will define national competitiveness for the next twenty years. [#SovereignAI](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fsovereignai&trk=public_post-text) [#DigitalSovereignty](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fdigitalsovereignty&trk=public_post-text) [#Geopolitics](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Fgeopolitics&trk=public_post-text) [#InfrastructurePower](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Finfrastructurepower&trk=public_post-text) [#AIControl](https://www.linkedin.com/signup/cold-join?session_redirect=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fhashtag%2Faicontrol&trk=public_post-text) [Alex Issakova](https://uk.linkedin.com/in/alexissakova?trk=public_post_reshare_feed-actor-name) Founder, Huckr AI | AI enablement & leadership advisory for B2B commercial teams | Ex-Sales Chief of Staff, Liftoff (pre-IPO, now Nasdaq: LFTO) | In AI since 2013 | Keynote speaker Why Are Countries So Obsessed With Sovereign AI? “Sovereign AI” has suddenly become a national priority for governments everywhere. Not because it sounds impressive, but because something deeper is shifting: AI is becoming infrastructure. And infrastructure is power. Here is what sovereign AI truly means, and why it matters. Before we go further, I’m writing a deeper series on this shift for my Substack The Long Signal — where I break down the architectures, the geopolitics, and what comes after the LLM bubble. If you want the full analysis as it unfolds, you can subscribe here: 👉 [https://lnkd.in/eqE3NuGH](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FeqE3NuGH&urlhash=VLqR&trk=public_post_reshare-text) \*\*1. Data Sovereignty Who shapes your AI’s worldview?\*\* Models absorb the values and assumptions of the data they are trained on. If that data comes from the Western internet, the worldview does too. Most countries lack enough local text to train large models from scratch, which means they inherit someone else’s cultural logic. Sovereign AI begins with sovereign data. \*\*2. Model Architecture Sovereignty Whose thinking style does your model inherit?\*\* Using Llama, Gemma or a Chinese-origin model means adopting the reasoning patterns, safety defaults and inductive biases of the creator. Fine-tuning adds local habits, but the underlying “mind” remains foreign. \*\*3. Compute Sovereignty Who controls your ability to train?\*\* Training needs extreme compute: NVIDIA GPUs, CUDA, PyTorch and more—mostly US-origin. Access can be restricted anytime. Without compute sovereignty, nations can’t independently train or scale models. \*\*4. Inference Sovereignty Where does national data actually run?\*\* If inference happens on foreign cloud systems, the data falls under foreign jurisdiction—impacting healthcare, policing, courts and national operations. \*\*5. Data Centre Sovereignty Where does your AI physically live?\* AI lives in buildings that need power, cooling, land and security. If those facilities are foreign-owned or located abroad, the country loses physical control over critical infrastructure. Data centres are the territorial layer of sovereign AI. \*\*6. AI Economy Sovereignty Who captures the value?\*\* A country can build a model yet lose everything if the value flows to foreign clouds, foreign vendors or foreign talent. True sovereignty requires a domestic AI economy that keeps and scales the benefits. The real reason sovereign AI matters AI is becoming the next foundation of national competitiveness. Productivity, industry, defence, culture and economic power will all be shaped by it. Countries are not obsessed with sovereign AI because it is trendy. They are obsessed because it decides who builds the future. ♻️ If this resonated, share it. Someone in your network needs this reminder today.  🔔 Follow [Alex Issakova](https://uk.linkedin.com/in/alexissakova?trk=public_post_reshare-text) for more reflections on the power dynamics shaping AI. 🧠 I have just launched my new Substack. The Long Signal. 👉 Subscribe here [https://lnkd.in/eqE3NuGH](https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Flnkd%2Ein%2FeqE3NuGH&urlhash=VLqR&trk=public_post_reshare-text)