Vibe Coding Hits the Navy and Dictionary as Security Flaws Mount

From meme to operating model

Vibe coding means describing the software you want in plain language and letting an AI agent write it. The idea started as a joke and now gets real budgets. In mid-September, a U.S. Navy acquisition office ran a two-day vibe coding workshop at Naval Base Point Loma in San Diego. IT staff and senior leaders there used agentic tools to prototype fixes for real Navy problems.22 Program officials said they want to roll the capability out across the enterprise. One assistant program manager went further: he wants junior Sailors and Marines to be able to build missing tools "in days instead of years" while deployed.22

The dictionaries have caught up too. Merriam-Webster added "vibe coding" in a batch of 1,400 new entries that also included "AGI" and a noun form of "compute."29 Commercial pressure is building at the same time. Lovable, a Swedish prompt-to-app company, raised $400 million in a Series C in August at a $13.3 billion valuation, twice the $6.6 billion it was worth in December.14

The overall picture is mixed. Prompt-driven app creation has become mainstream. Methods for checking that software have not kept up, and almost every serious study of the output says so.

The money: fast growth, priced for more of it

Lovable's numbers are the clearest example of the boom. The company says annual recurring revenue has almost tripled from $200 million and was on track for $600 million by the end of August.4 It also reports more than 60 million projects created since its November 2024 launch, and says apps built on Lovable get more than 900 million visits a month.69 The money will go toward enterprise integrations, governance and permissions controls, training Lovable's own models, and hiring. The company plans to reach about 450 staff by year's end, with most new hires in machine learning, infrastructure and security.1

The round also came in above earlier expectations. In June, reports had Lovable seeking about $12 billion.2 One analysis puts the final price at roughly 22 times forward revenue. That is expensive, but not out of line given how fast the company is growing.5 More cautious reporting points out that the revenue, project and traffic figures all come from the company and have not been independently verified. Project counts also include plenty of free or abandoned experiments that never turn into revenue.6 One industry report puts it bluntly: 50 million projects does not mean 50 million apps.28

Lovable is not the only company raising at this pace. India's Emergent tripled its valuation to about $300 million in January with a $70 million round led by SoftBank Vision Fund 2 and Khosla Ventures. It claimed $50 million in ARR and more than 5 million users.7 The biggest deal of all was SpaceX's all-stock purchase of Cursor's parent company, Anysphere, valued at $60 billion.25 Reports give different dates for that deal. One account says the agreement was reached in mid-June, while a market report dates the acquisition to August 14.224 Either way, the deal sets the price for AI coding tools. Investors are betting that writing software will open up to far more people, not just that existing engineers will get a bit faster.2

The product race: closing the loop from prompt to production

In the past few weeks, product launches have mostly targeted the steps that come after the first prompt. Replit released a Windows desktop app that runs AI-generated builds locally, with each one sandboxed.27 Bolt.new now lets reviewers click any element in a preview and leave a comment, which goes straight to the AI agent as a task.27 Lovable switched its app-building agent to Anthropic's Opus 5.5 and says builds now take noticeably fewer steps.27 Google Labs launched Playground, which turns text prompts into playable browser games.27

The idea is also spreading beyond developer tools. Alaio's Vibecode plugs into Bitrix24, so employees can describe tools like CRM lead scoring or tasks generated from meeting notes and have them built on top of their existing business data.26 Founder-focused coverage notes that "vibe deploying" is narrowing the gap between a prototype and a live product. Build, test and publish now increasingly happen in one flow.23

One recent benchmark complicates the industry's push toward multi-agent systems. Vals AI ran leading models on its Vibe Code Bench both as solo coders and as teams of agents. The teams cost about five times as much and rarely scored higher.27 In practice, this suggests better results will come from tighter human review rather than piling on more agents.

The security ledger: consistent, and not good

Reports disagree about adoption and valuations. On security they line up closely. In June, Tenzai tested Cursor, Claude Code, OpenAI Codex, Replit and Devin by having each build the same applications. It found 69 vulnerabilities across 15 apps, and every agent shipped vulnerable code.14 All five added a server-side request forgery flaw when asked to build a link-preview feature. None of the 15 apps had working cross-site request forgery protection.14

Research on live deployments points the same way. Escape scanned more than 5,600 publicly available vibe-coded apps. It found over 2,000 vulnerabilities, more than 400 exposed secrets and 175 cases of exposed personal data.1516 Veracode's 2026 testing found that about 44 percent of AI code generation tasks introduced a risky vulnerability. Its average security pass rate was 56 percent, almost unchanged from its first report.16 Georgia Tech's Vibe Security Radar tracks vulnerabilities that can be traced to AI coding tools. Its count of official cases rose from 6 in January 2026 to 35 in March. Its founder estimates the real number is five to ten times higher, because most AI-written code leaves no trace of where it came from.28

The Cloud Security Alliance highlights how the problem is changing. Apiiro data from Fortune 50 companies showed AI-assisted developers committing code three to four times faster. Syntax errors dropped, but privilege escalation paths rose 322 percent and architectural design flaws rose 153 percent.11 Simple mistakes are disappearing, while flaws that need an understanding of the whole system are increasing. Those are exactly the flaws a non-developer working from prompts is least able to spot.

The category's best-known incident shows how this happens. Researchers found that 170 of 1,645 Lovable-built apps had flaws that let anyone access personal information.21 The security alliance says controls like row-level database access policies are where AI code generators perform worst. These controls require deliberate, system-wide intent.19

Productivity: contested, and getting harder to measure

The productivity question is less settled than the boosters suggest. Sonar's 2026 survey found that 96 percent of developers do not fully trust AI-generated code to be functionally correct. Asked about the future, respondents ranked reviewing and validating AI output as the most important skill, ahead of prompting.13 The same survey shows vibe coding is most effective for greenfield prototyping. AI scores much worse at changing existing code, debugging and refactoring.13 Stack Overflow data shows that 66 percent of developers are frustrated by AI output that is "almost right, but not quite." Another 45 percent say debugging it can take longer than writing the code themselves.28

METR's controlled research is a revealing signal. When the team repeated its study in 2026, developers increasingly refused tasks that had to be done without AI. That broke the experimental design, and the researchers now treat their newer estimate as only a lower bound.28 In other words, developers won't give up the tools, even while the evidence that the tools make them faster is still incomplete.

The pushback is coming from respected engineers and from communities that rely on careful work. Two engineers behind the Pi coding harness have warned of a coming "vibe slop" crisis. In their view, companies are trading near-term speed for outages, security holes and technical debt.21 GitHub has acknowledged that low-quality AI-generated contributions are overwhelming open-source maintainers. cURL, for example, ended its bug bounty after AI-generated reports multiplied.21 Fan developers who port retro games are split as well. A lead contributor to the Banjo-Kazooie PC port argues that AI-built ports often stall because their makers don't understand the original games well enough.25 A developer who used Claude Opus 5.5 and Rust to recreate seven Adobe applications showed how much is now possible. He also admitted the result is an early alpha and "not anywhere close to ready."30

The reading: create cheaply, verify expensively

Put together, the reporting points to one conclusion. Vibe coding has settled the question of whether ordinary people can create software. It has not settled whether that software can be trusted. One industry report sums it up: vibe coding is a mass-market way to create software and a minority way to ship it. That report also found that 60.5 percent of builders surveyed on one major platform were not yet making money from their projects.28

The valuable work is moving to verification. Lovable's hiring plan leans toward security and governance, Bolt is building review into its product, and the Navy wants agentic coding inside its development and security pipeline.12722 Even the person who coined the term has moved on. In February, Andrej Karpathy called vibe coding "passe" and proposed "agentic engineering," where agents write the code and humans own the architecture and review.24

Investors are pricing in a world where anyone can build software. The security data says that world will also produce a lot of software nobody has properly checked. The tools that can show their output is safe will be the ones that justify today's valuations.