Sold as finished: the vibe coding (trust) collapse

On 2 February 2025, a researcher who had spent his career training some of the most capable software models in the world posted a short definition that would eventually earn its own dictionary entry. "There's a new kind of coding I call 'vibe coding'," Andrej Karpathy wrote, "where you fully give in to the vibes, embrace exponentials, and forget that the code even exists."¹ He was describing his own hobby projects: disposable, low-stakes, forgiving of mistakes.

It mostly worked for Karpathy because Karpathy already knew what "mostly" was hiding. A year later, the phrase had become the default onboarding experience for software with a database, a login page, and other people's data sitting behind it — and the people doing the vibing rarely had that context to fall back on. In early 2026, a social network for AI agents called Moltbook launched to considerable online fanfare, built by a founder, known publicly only by the handle @mattprd, who described the process on X: "I didn't write a single line of code for @moltbook. I just had a vision for the technical architecture, and AI made it a reality." On 31 January 2026, researchers at the security firm Wiz found a publishable Supabase key sitting inside the client-side JavaScript bundle. Because nobody had ever turned on Row Level Security, that key granted full read and write access to the production database. Exposed: 1.5 million agent API authentication tokens, 35,000 user email addresses, and 4,060 private agent-to-agent conversations — some containing plaintext OpenAI API keys. Moltbook's team shipped emergency patches within hours of disclosure.²

Nobody broke into Moltbook. Nobody had to. The door was open because nobody had thought to check whether it was closed, and the person who built the house did not know houses had doors that needed checking.

I keep circling back to that story, and not because @mattprd did anything unusually foolish. He did exactly what he was told he could do. Somewhere between the marketing copy and the model card, an entire industry taught him that this was fine.

How software got broken

Twenty-six years ago a consultant named Mark Minasi noticed something the rest of manufacturing had never quite managed: you could ship a product with defects, disclaim the warranty in the licence agreement nobody reads, and the market would absorb it anyway.³ Cory Doctorow later gave the mature version of that pattern a name everyone now uses without remembering where it came from.⁴ I've made this argument at length before, so I won't re-run it here — the short version is that we stopped buying software and started renting access to it, stopped expecting a finished product and started accepting a permanent beta with a changelog, and somewhere in that transition the word "update" quietly lost its promise. A 2026 study found Disney+'s ad-free tier had risen 116.9% since 2020 — the largest inflation-adjusted price increase of any major subscription service.⁵ Nobody voted on any of this. Nobody was handed a ballot. There was a notification, then another notification, then a feature that used to be included quietly reappeared behind a higher tier with a friendlier name, and somewhere in the accumulation of small, reasonable-sounding asks, we learned that objecting was more effort than it was worth. We just kept clicking "accept" often enough that the industry stopped bothering to ask.

This is the part that actually matters for everything that follows: a market that spent two decades training its customers not to expect quality was exactly the market a half-finished technology needed in order to look normal. An assistant that hallucinates, hedges, and states things confidently that are not true would have been a scandal in 2005. In 2026 it is a subscription tier, and the tier is selling well.

Say plainly what is actually being sold. A large language model marketed as a coding assistant is not a finished product with a few rough edges still to sand down. It is an incomplete one — its failure domains are known well enough to be written up in the model cards and safety evaluations vendors themselves publish, and largely absent from the marketing copy that reaches a citizen developer deciding whether to trust it with a signup form, a payment flow, or a database. The US Securities and Exchange Commission has already built a docket around exactly this asymmetry, under the label "AI washing": in 2024 it fined the investment adviser Delphia $225,000 and Global Predictions $175,000 for claiming to run machine-learning-driven portfolios that, on inspection, used no such technology at all.⁶ Tesla sold nine years of full self-driving "next year" before its own chief executive admitted, on a call in January 2025, that cars with Hardware 3 — the computer Tesla had shipped since 2019, expected to deliver on a promise first made in October 2016 — would need a physical hardware upgrade before they could run the unsupervised full self-driving that had been promised of them.⁷ The pattern is not confined to finance and cars. New Relic's 2026 survey of 200 US technology decision-makers found 94% rated AI-generated code as higher quality than human-written code at the moment it was reviewed — and 82% of that same group had suffered at least one major production failure caused by AI-generated code within the past six months, with 78% reporting a measurable spike in incidents directly tied to it.⁸

That is the mechanism underneath everything in this piece. A bad product — immature, imperfectly understood even by the people who built it, sold as more capable and more finished than it is — has become the primary tool handed to the least prepared population in software's history, and is now being used to manufacture more bad products in its own image. Vibe coding did not invent a new kind of mistake. It industrialised an old one, and handed it to a generation of builders who were never warned the tool itself was still half-built.

The scale nobody priced in

Gartner's forecast put the shift in blunt terms: over 70% of new applications developed by enterprises will use low-code or no-code technologies by 2025.⁹ A 2026 survey found 63% of vibe-coding and AI-app-builder users report no coding background whatsoever — not junior, not self-taught, none.¹⁰ Industry analysis puts developers outside formal IT departments on track to account for at least 80% of the user base for low-code tools by 2026, and large enterprises are already estimated to run roughly four citizen developers for every professional one.¹¹

None of this makes the people doing the building reckless. Most are doing precisely what the tools were marketed to let them do: describe an idea in plain language and watch it become a working product, sold by an industry with every incentive to let them believe that description was the hard part.

Nobody who can tell you the ground is unsteady

It would be tidy if this were only a story about people who never learned to code. It isn't. The Cloud Native Computing Foundation's own landscape — the closest thing this industry has to an inventory of what "building software properly" now requires — lists well over a thousand tools and products across provisioning, orchestration, observability, security, and delivery.¹² Most working engineers use a handful of them and have made peace with not knowing the rest. That peace was survivable when a team had specialists for each layer and a senior engineer whose entire job was knowing which corners could and couldn't be cut. It stops being survivable the moment that senior engineer becomes a team of one, reaching for an AI assistant to cover the layers they never had to learn, because someone always covered them before.

A model that writes secure Python and dangerous Java does not announce the difference; it writes both with identical, cheerful confidence. A frontend specialist who has never had reason to think about database access control leans on an assistant to stand up a backend for a side project, and inherits whatever blind spot that model happens to carry in that domain, with no personal experience to flag that anything is wrong, because nothing about the output looks wrong. A junior engineer inside a real organisation has this same blind spot on their first day, and it rarely matters, because a senior colleague reviews the pull request, a scanner runs in the pipeline, and a deploy gate exists to catch the mistake before production. Vibe coding removes the junior engineer's inexperience from the equation without replacing any of the scaffolding that used to make that inexperience survivable.

I don't think the solo senior engineer's version of this is a smaller problem than the citizen developer's. I think it is the same problem wearing a better résumé. There is a particular kind of confidence that shows up in a senior engineer standing alone in front of a service they have never operated before, watching an assistant scaffold an authentication flow in seconds and thinking, reasonably, that ten years of judgement should be enough to catch whatever the tool gets wrong. Ten years of judgement built inside one specialism does not automatically transfer to a different one, and the tool gives no signal that it has crossed into territory the judgement doesn't cover. A C-suite executive signing off on a vendor stack, a solo engineer standing up infrastructure they've never operated, and a founder who has never written a line of code are all making the identical bet: that the tool understands the terrain better than they do. The data on how often that bet quietly fails is not a coding-specific curiosity. It is the whole shape of this piece, arriving from a different direction.

That bet is not evenly distributed, though, and it is worth being honest about why. A solo engineer or a citizen developer has no particular reason to want the tool to look more capable than it is; their only real incentive is that the thing works. An executive deciding whether to adopt an AI vendor, a platform, or an entire AI-first workflow is carrying a second incentive alongside the first, and it pulls the opposite way: the tool being capable enough looks identical, on a slide in a board meeting, to the tool actually being capable enough, and only one of those requires the slower, more expensive, less career-flattering path of properly checking. Deloitte's Q2 2026 survey of 200 North American finance chiefs at companies with revenues over $1 billion found 59% cite balancing pressure to deploy AI quickly against managing the risk of doing so as their single biggest governance challenge — and 96% expressed confidence in their AI governance frameworks regardless, which is the kind of number that should make a reader more nervous, not less.¹³ Adoption, in other words, is rarely being driven by evidence that a tool works. It is being driven by the promise that it might, priced in and acted on long before the proof arrives.

Ask how often that promise gets checked before the cheque clears. In January 2025 the SEC settled with Presto Automation, a restaurant technology company that had marketed its "Presto Voice" drive-through ordering system as an in-house AI capable of eliminating human order-taking entirely. Until September 2022, the AI actually running the system was built, owned, and operated by an undisclosed outside supplier — and even after Presto built its own proprietary version, that system still required substantial human involvement.¹⁴ Nobody buying that system for a chain of restaurants independently verified the number. They bought the label.

The mistake that keeps repeating

Veracode's 2026 GenAI Code Security Report tracked over 100 models across four testing snapshots and found the average security pass rate for AI-generated code sits at 56%, essentially unchanged across four years of tracking despite the volume of AI-written code surging in the meantime.¹⁵ Java fared worst of all tested languages at a 30% pass rate; the best-performing model of the year, GPT-5.5, still only reached 68%.¹⁵ Separately, Veracode's testing found 45% of AI-generated code samples introduced OWASP Top 10 vulnerabilities, with 86% failing to defend against cross-site scripting and 88% vulnerable to log injection, while enterprise commit data showed privilege-escalation paths rising 322% and architectural design flaws rising 153% among AI-assisted changes.¹⁶

Georgia Tech's Systems Software & Security Lab runs a public tracker called Vibe Security Radar for precisely this reason: by April 2026 it had confirmed 74 CVEs directly attributable to AI-generated code — 14 critical, 25 high-severity — and March 2026 alone produced 35 of those cases, more than the 18 the lab found across the entire second half of 2025.¹⁷ Auditors who specialise in this category of software report that roughly 70% of the Lovable-built applications they examine have row-level security disabled entirely — not misconfigured, simply never switched on — with a typical build carrying eight to fourteen distinct security findings before anyone outside the founder has ever looked at it.¹⁸ The mechanism is consistent: the assistant builds the feature that was asked for, and configures security only when the prompt explicitly asks for it. Nobody asks for it, because nobody who has never had a database breached knows that asking is a step that exists. You cannot ask a question you don't know exists to be asked, and no amount of enthusiasm closes that particular gap.

The models have a second, stranger failure mode that has nothing to do with what a human forgot to specify. Researchers generated 576,000 code samples across sixteen popular models and found hallucination rates of at least 5.2% for commercial models and 21.7% for open-source ones — plausible-sounding package names that were never published anywhere, more than 205,000 unique fictional names across the study.¹⁹ A 2026 re-evaluation found the per-model rate had fallen, but identified 127 package names that five different models all hallucinated identically, of which 53 remained available for registration as of April 2026 — meaning an attacker could pre-register more than four in ten of them and simply wait for a model to keep recommending its own fabrication.^{20 21} Attackers do not need to breach anything here. They just need to register the lie before anyone else notices it was one.

Scale is what turns a pattern of small incidents into an industry problem, and the clearest recent illustration did not involve a hobby project at all. In 2025 the AI music platform Suno disclosed a data breach that exposed more than 55 million email addresses and tens of thousands of Stripe payment records containing partial credit card data.²² Suno is a funded, staffed, AI-native company. The breach shows exactly what the Moltbook pattern looks like once it reaches the financial data that vibe-coded applications are already collecting by default, at a scale that makes the overlap between "unaccountable" and "financially catastrophic" look less like speculation and more like a matter of when.

Nobody to sue

The largest GDPR fine ever issued — €1.2 billion, imposed on Meta Ireland in May 2023 for unlawful international data transfers — represented a genuine record, and it was still a figure the company's balance sheet absorbed without much visible strain.²³ A €9.55 million fine against another company was cut by 90% on appeal after a court found the amount disproportionate.²⁴ IBM's 2026 Cost of a Data Breach Report put the global average breach cost at $4.99 million, up 12% on the previous year and a record high — real money, but money that only a company already operating at that revenue scale could ever be made to pay, and rarely money that ends anyone's career.²⁵ Regulators have started shifting toward personal liability for executives specifically because fines alone were not deterring anyone; the Dutch Data Protection Authority's investigation into whether Clearview AI's own directors can be held personally liable is a tacit admission that the old model failed.²⁴ Grand visions, it turns out, are usually what people call the period just before the invoice arrives — the invoice just hadn't found Meta yet, and when it did, Meta kept operating.

A company that size absorbs a breach through machinery built years in advance for exactly this purpose: cyber insurance underwritten long before the incident, an outside forensics firm on retainer, in-house counsel who already know which regulator to call first. None of that machinery exists behind a Supabase project spun up on a Tuesday afternoon, and that is the less interesting half of the argument, because it implies the other half is the opposite. It isn't. Every US state requires breach notification regardless of company size, and California allows statutory damages of $100 to $750 per affected person, per incident, without the plaintiff needing to prove any actual financial loss.²⁶ On paper, the solo operator carries the same exposure as the enterprise. In practice, the exposure is uncollectable. There are no assets behind a limited liability company capitalised with a laptop and a subscription. The business can be dissolved in an afternoon, well before a plaintiff's letter finds a forwarding address. Limited liability companies are optimised for exactly one kind of limiting, and it is not the kind aimed at the people who signed up trusting the login page.

George Akerlof described the economic consequence of exactly this asymmetry in 1970, using used cars rather than software.²⁷ When a buyer cannot verify the quality of a good before purchasing it, sellers of genuinely good products cannot charge what that quality is worth, because buyers have no reliable way to distinguish it from a bad one dressed identically. The average quality of what remains on the market falls, buyers trust the market even less, and the cycle compounds. Software buyers cannot inspect a codebase before handing over a date of birth or a card number. The signals available to them — a polished landing page, a slick onboarding flow, a confident-sounding privacy policy — are exactly the signals AI-assisted builders are best at producing, entirely independent of what security setting sits behind the interface. The market has invented trust signals before: SOC 2 reports, penetration-test attestations, certifications that exist precisely because Akerlof's problem is old news to enterprise buyers who learned decades ago that a vendor's word is not evidence. Those signals assume a vendor with a compliance budget and a sales cycle long enough to justify the audit. None of that describes an app that went from prompt to production in an afternoon, which leaves buyers exactly where Akerlof's second-hand car shoppers were: guessing from the paint job, in a market where the paint job has never been cheaper or easier to fake.

You cannot opt out

It would be reasonable, at this point, to conclude none of this touches you. You don't use AI app builders, you've never signed up for a hobby project built by a stranger on the internet, and you are nowhere near a vendor stack decision. That conclusion assumes the risk only travels through doors you personally chose to walk through, and it doesn't.

In June 2025, security researchers found that McHire, the recruitment platform used by McDonald's — one of the largest employers on the planet — ran its hiring process through a chatbot called Olivia, built and operated by a company called Paradox.ai. Olivia's administrative backend was protected by nothing more than a test account that had been left active since 2019, secured with the password "123456." Sixty-four million job applications were exposed: names, contact details, employment history, authentication tokens, and chat transcripts.²⁸ Nobody who applied for a McDonald's job chose Paradox.ai; the choice was never on offer to them. They chose McDonald's — a name about as far from "vibe-coded startup" as software gets — and the risk found them anyway, two or three vendor relationships removed from anything they actually decided.

That is the part "just don't use risky software" misses entirely. You do not need to personally adopt an AI-built tool to inherit its failure. You need only use a product, however established, that has quietly adopted one somewhere upstream of you — in its recruitment pipeline, its customer support widget, its payments backend, its analytics stack — usually without disclosing that it did, and rarely because anyone checked whether the tool deserved the trust it was handed. Paradox.ai is not a hobbyist's weekend build. It had raised $200 million in a single funding round at a $1.5 billion valuation, and counts Unilever, Nestlé, CVS Health, General Motors, and McDonald's among its 500-plus named clients — a genuinely established vendor by any reasonable definition — and it still left a live administrative door unlocked for six years.²⁹ Multiply that same gap by the millions of far smaller, far less funded, far more recently assembled AI tools now quietly integrated into the products you already trust, and "I don't use vibe-coded software" stops being a meaningful defence. It was never really available as one.

Only one lever has ever actually worked

I could end here with a list of things platforms could do, and I largely already have, elsewhere in pieces about this same territory: Row Level Security on by default, disclosure standards for known failure modes, insurers pricing the aggregate risk instead of chasing individual builders. All of it is real and all of it would help. None of it is a strategy, because none of it is something this industry has ever done voluntarily at the pace the harm accumulates. The SEC didn't stop AI washing by asking nicely; it had to fine people. Tesla didn't admit Hardware 3 couldn't deliver what it promised out of good faith; it took nine years and a shareholder call. An industry that has spent two decades learning that broken ships faster and cheaper than correct does not self-correct because a blog post asked it to.

There is precedent for what does work, and it comes from an industry with a worse history than software's, not a better one. Before 1962, US pharmaceutical firms did not have to prove a drug was effective before selling it, only that it wasn't obviously dangerous by the standards of the time — a standard thalidomide slipped past in much of Western Europe, causing thousands of severe birth defects before its effects were understood.³⁰ It took the Kefauver-Harris Amendment, passed by Congress in October 1962 in direct response, to require firms to prove both safety and effectiveness before marketing a drug at all.³⁰ The industry did not arrive at that standard by choice. It was handed the standard, at cost, after the alternative had already hurt people. Software has been allowed to disclaim liability for hallucinating for thirty years because for most of that time, when it broke, the worst case was an afternoon of frustration. That stopped being true a while ago, and the only lever that has ever closed a gap like this one is regulation an industry is forced to answer to, not invited to consider.

I want to be honest about where the discourse actually is, because it is more sober than the loudest voices in it suggest. An academic study tracking practitioner discussion of vibe coding through 2026 found the conversation maturing rather than curdling — moving from early optimism, through a mid-year stretch of scepticism and fatigue, toward something closer to responsible integration and shared ownership of the risk.³¹ Software people are not oblivious to any of this. They are, on the whole, not the ones with the authority to force the industry's hand.

Australia already ran a version of this experiment, on hard mode, with companies that had security teams, budgets, and legal departments. In September 2022 a coding error introduced in 2018 left a customer-data API open to the internet without a credential check, and the Australian Information Commissioner alleges that from October 2019 to September 2022 this exposed the personal information — names, dates of birth, addresses, passport and licence numbers — of roughly 9.5 million Optus customers, over a third of the country. That case is still before the Federal Court.³² Two months after the Optus breach became public, Medibank confirmed ransomware attackers had accessed the records of 9.7 million customers, including 480,000 health claims, and had published sensitive treatment data — including for mental health and chronic illness — after Medibank declined to pay. The regulator alleges Medibank knew of serious cyber security deficiencies, including a lack of multi-factor authentication, for at least eighteen months before the attack. That case, too, remains before the Federal Court, years later.³³

Almost twenty million Australians, between the two, at companies with full-time engineers and dedicated security budgets. The failure that let it happen was, underneath the headlines, a small and boring access-control mistake — the same category of mistake this entire piece has spent most of its length describing. That was the accountable version of this failure, and it is still working its way through court years afterward. What's arriving now is the unaccountable version, built by orders of magnitude more people, on tools that are half-finished by design, at a pace no regulator currently tracks in anything close to real time. It will not look like Optus. It will look like thousands of smaller ones nobody outside the affected users ever hears about, until enough of them land on the same week that someone finally has to count.

I keep returning to that detail from Moltbook: a publishable key sitting in a client bundle, a security setting that was never so much as a checkbox in the builder's mental model, a founder who genuinely believed a vision for the architecture was the same thing as the architecture itself. He was not lying. He simply could not see the part of the house that was never built, because nobody who sold him the tools had any incentive to point it out. Multiply that blind spot by the millions of citizen developers already building, the solo engineers quietly doing the same thing with better titles, and the millions more arriving each year. Subtract any realistic path to accountability when it fails. You do not get a story about bad actors. You get a market nobody can verify, filling up with software nobody can vouch for, priced and marketed identically to the software that actually deserves your date of birth — built, at every layer, on tools an industry marketed as finished while its own model cards, safety evaluations, and regulators' enforcement dockets said otherwise.

---

Footnotes

1. Karpathy, A. (2025). 'There's a new kind of coding I call "vibe coding"...' X (Twitter). https://x.com/karpathy/status/1886192184808149383 ↩

2. Wiz. (2026). 'Hacking Moltbook: AI Social Network Reveals 1.5M API Keys.' Wiz Blog. https://www.wiz.io/blog/exposed-moltbook-database-reveals-millions-of-api-keys ↩

3. Minasi, M. (2000). 'The Software Conspiracy: Why Companies Put Out Faulty Software, How They Can Hurt You and What You Can Do About It.' McGraw-Hill. ↩

4. Doctorow, C. (2023). 'Tiktok's enshittification.' Pluralistic. https://pluralistic.net/2023/01/21/potemkin-ai/ ↩

5. DepositAccounts. (2026). 'Paying More, Using Less: Digital Subscription Prices Climb Since 2020 as Consumers Cancel and Reconsider.' DepositAccounts. https://www.depositaccounts.com/blog/digital-subscription-inflation-study.html ↩

6. U.S. Securities and Exchange Commission. (2024). 'SEC Charges Two Investment Advisers with Making False and Misleading Statements About Their Use of Artificial Intelligence.' SEC.gov. https://www.sec.gov/newsroom/press-releases/2024-36 ↩

7. TechCrunch. (2025). 'Elon Musk reveals Elon Musk was wrong about Full Self-Driving.' TechCrunch. https://techcrunch.com/2025/01/30/elon-musk-reveals-elon-musk-was-wrong-about-full-self-driving/ ↩

8. New Relic. (2026). 'The 2026 State of AI Coding Report.' New Relic. https://newrelic.com/resources/report/2026-state-of-ai-coding ↩

9. ToolJet. (2026). 'Gartner Forecast on Enterprise Low-Code Development Technologies in 2026.' ToolJet Blog. https://blog.tooljet.com/gartner-forecast-on-low-code-development-technologies/ ↩

10. TechTimes. (2026). 'Vibe Coding for Non-Developers: 63% of Users Now Have No Coding Background, Breaches Follow.' TechTimes. https://www.techtimes.com/articles/317077/20260524/vibe-coding-non-developers-63-users-now-have-no-coding-background-breaches-follow.htm ↩

11. Kissflow. (2026). 'Citizen Development Trends & Key Stats 2026.' Kissflow. https://kissflow.com/citizen-development/citizen-development-statistics-and-trends/ ↩

12. CNCF. (2026). 'Cloud Native Landscape.' Cloud Native Computing Foundation. https://landscape.cncf.io/guide ↩

13. PR Newswire / Deloitte. (2026). 'CFOs Face Pressure to Deploy AI Quickly While Managing Risk is a Top Challenge from Latest Deloitte Q2 2026 CFO Signals Survey.' PR Newswire. https://www.prnewswire.com/news-releases/cfos-face-pressure-to-deploy-ai-quickly-while-managing-risk-is-a-top-challenge-from-latest-deloitte-q2-2026-cfo-signals-survey-302832785.html ↩

14. Cooley PubCo. (2025). 'SEC charges "AI-washing" at Presto Automation.' Cooley LLP. https://cooleypubco.com/2025/01/30/sec-charges-ai-washing/ ↩

15. Veracode. (2026). '2026 GenAI Code Security Report: 100+ Models Tested.' Veracode. https://www.veracode.com/resources/analyst-reports/2026-genai-code-security-report/ ↩ ↩²

16. Veracode. (2026). 'AI Coding Tools Are Creating a Security Gap We Must Close Immediately.' Veracode Blog. https://www.veracode.com/blog/ai-coding-tools-security-gaps/ ↩

17. Georgia Tech Research. (2026). 'Bad Vibes: AI-Generated Code is Vulnerable, Researchers Warn.' Georgia Institute of Technology. https://research.gatech.edu/bad-vibes-ai-generated-code-vulnerable-researchers-warn ↩

18. Vibe Coding Academy. (2026). 'Vibe Code Audit: Secure Your AI App Before Launch.' vibecoding.app. https://vibecoding.app/blog/vibe-code-audit ↩

19. Spracklen, J., Wijewickrama, R., Sakib, A. H. M. N., Maiti, A., Viswanath, B., & Jadliwala, M. (2025). 'We Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs.' 34th USENIX Security Symposium. https://www.usenix.org/conference/usenixsecurity25/presentation/spracklen ↩

20. 'The Range Shrinks, the Threat Remains: Re-evaluating LLM Package Hallucinations on the 2026 Frontier-Model Cohort.' (2026). arXiv. https://arxiv.org/pdf/2605.17062 ↩

21. CSO Online. (2026). 'Top AIs invent same fake PyPI and npm package names.' CSO Online. https://www.csoonline.com/article/4201164/top-ais-invent-same-fake-pypl-and-npm-package-names-2.html ↩

22. TechNadu. (2026). 'Suno Data Breach Exposes 55M Email Addresses and Stripe Records.' TechNadu. https://www.technadu.com/suno-data-breach-55-million-emails-and-stripe-records-exposed/631566/ ↩

23. Faegre Drinker. (2023). 'Meta Fined EUR 1.2 Billion for Violating GDPR.' Faegre Drinker Biddle & Reath LLP. https://www.faegredrinker.com/en/insights/publications/2023/5/meta-fined-eur-1-2-billion-for-violating-gdpr ↩

24. Infosecurity Magazine. (2026). 'Deterring Data Privacy Violations in Big Tech: Why Fines Aren't Enough.' Infosecurity Magazine. https://www.infosecurity-magazine.com/news-features/data-privacy-violations-big-tech/ ↩ ↩²

25. IBM. (2026). 'Cost of a Data Breach Report 2026.' IBM. https://www.ibm.com/reports/data-breach ↩

26. Turley Law. (2026). 'Can You Sue a Company for a Data Breach? What You Need to Know.' Turley Law. https://turleylaw.com/blog/sue-company-data-breach ↩

27. Akerlof, G. A. (1970). 'The Market for "Lemons": Quality Uncertainty and the Market Mechanism.' The Quarterly Journal of Economics, 84(3), 488–500. https://doi.org/10.2307/1879431 ↩

28. Krebs, B. (2025). 'Poor Passwords Tattle on AI Hiring Bot Maker Paradox.ai.' Krebs on Security. https://krebsonsecurity.com/2025/07/poor-passwords-tattle-on-ai-hiring-bot-maker-paradox-ai/ ↩

29. PR Newswire. (2022). 'On a Mission to Build the Next Generation of HR and Recruiting Software through Conversational AI, Paradox Raises $200M Series C.' PR Newswire. https://www.prnewswire.com/news-releases/on-a-mission-to-build-the-next-generation-of-hr-and-recruiting-software-through-conversational-ai-paradox-raises-200m-series-c-301450591.html ↩

30. U.S. Food and Drug Administration. (2006). 'Promoting Safe & Effective Drugs for 100 Years.' FDA. https://www.fda.gov/about-fda/histories-product-regulation/promoting-safe-effective-drugs-100-years ↩ ↩²

31. 'Hope or Hype? Understanding Vibe Coding through Software Practitioner Discussions.' (2026). Proceedings of the 19th International Conference on Cooperative and Human Aspects of Software Engineering. https://dl.acm.org/doi/10.1145/3794860.3794904 ↩

32. Office of the Australian Information Commissioner. (2025). 'Australian Information Commissioner takes civil penalty action against Optus.' OAIC. https://www.oaic.gov.au/news/media-centre/australian-information-commissioner-takes-civil-penalty-action-against-optus ↩

33. Office of the Australian Information Commissioner. (2024). 'OAIC takes civil penalty action against Medibank.' OAIC. https://www.oaic.gov.au/news/media-centre/oaic-takes-civil-penalty-action-against-medibank ↩