Vibe coding statistics (2026)

Data loss and production incidents

Cases where AI coding agents deleted data or broke production, how each one ended, and Anthropic's own numbers on catching risky commands.

• 1,206 records

  1,206 executive records and 1,196+ company records: that's what Replit's AI agent said it destroyed when it ran a database push without permission during a code freeze on SaaStr founder Jason Lemkin's app (July 2025). Hours later, Lemkin restored the data with Replit's rollback.

  Source: Jason Lemkin on X, July 18, 2025. Also: Jason Lemkin on X, July 18, 2025

  The record counts are the agent's own claim, not verified numbers.

  Copy citation
• 2 days

  2 days after Lemkin's post, Replit CEO Amjad Masad called the deletion "unacceptable and should never be possible," said Replit was rolling out automatic dev/prod database separation, and noted that backups already allowed a one-click restore of the whole project.

  Source: Amjad Masad on X, July 20, 2025

  Copy citation
• 4,000 records

  4,000 records of fictional people: Lemkin said in a LinkedIn video that Replit's agent had also filled a database with made-up data (via The Register).

  Source: The Register, “Vibe coding service Replit deleted production database, faked data, told fibs galore”, July 21, 2025

  The Register's summary of a LinkedIn video, not Lemkin's own words.

  Copy citation
• 9 seconds

  9 seconds: that's how long PocketOS's founder says it took a Cursor agent running Claude Opus 4.6 to delete the company's production database, along with the backups it could see, through one Railway API call (April 24, 2026). Railway restored the data two days later.

  Source: Jer Crane (PocketOS founder) on X, “An AI Agent Just Destroyed Our Production Data. It Confessed in Writing.”, April 25, 2026. Also: Railway, “Your AI wants to nuke your database”, April 29, 2026. Also: ABC7 News, “'Rogue' AI agent from SF-based Cursor goes haywire, deletes company's entire database”, April 30, 2026

  Railway says it had separate offsite disaster backups and restored the data from them.

  Copy citation
• 3 months

  3 months: PocketOS's newest backup it could restore on its own was three months old, so its rental-company customers spent the weekend rebuilding bookings from Stripe, calendars and email, until Railway recovered the full database two days after the deletion.

  Source: Jer Crane (PocketOS founder) on X, “An AI Agent Just Destroyed Our Production Data. It Confessed in Writing.”, April 25, 2026. Also: Railway, “Your AI wants to nuke your database”, April 29, 2026. Also: ABC7 News, “'Rogue' AI agent from SF-based Cursor goes haywire, deletes company's entire database”, April 30, 2026

  The founder's post was written more than 30 hours after the deletion, before the data was recovered.

  Copy citation
• 30 minutes

  30 minutes: once Railway's CEO got involved, his team restored PocketOS's data within half an hour, Railway's CEO told ABC News and the founder said on X, about two days after the deletion. Railway has since made every delete through its API a 48-hour soft delete.

  Source: ABC7 News, “'Rogue' AI agent from SF-based Cursor goes haywire, deletes company's entire database”, April 30, 2026. Also: Railway, “Your AI wants to nuke your database”, April 29, 2026

  The 30 minutes counts from when Railway's CEO connected with the founder, not from the deletion.

  Copy citation
• Jul 21, 2025

  July 21, 2025: Gemini CLI (v0.1.13, Gemini 2.5 Pro) told a Windows user it had lost every file in his project folder after a botched move, saying "I have failed you completely and catastrophically." The bug report doesn't say whether the files were ever recovered.

  Source: GitHub issue #4586 in google-gemini/gemini-cli, “Gemini CLI 'lost' files during a failed file move operation. [Windows]”, July 21, 2025

  The reporter wrote that he expected the tool not to "hallucinate 'losing' the files."

  Copy citation
• 13 hours

  13 hours: that's how long an AWS cost-management service was down in December 2025 after engineers let Amazon's Kiro coding agent make changes and it chose to "delete and recreate the environment," the Financial Times reported, citing unnamed sources. Amazon says the cause was a misconfigured role, not AI, and that only AWS Cost Explorer in one of its 39 regions was affected.

  Source: Reuters (citing the Financial Times), “Amazon's cloud unit hit by outage involving AI tools in December”, February 20, 2026. Also: Amazon, “Correcting the Financial Times report about AWS, Kiro, and AI”, February 20, 2026

  Copy citation
• Feb 2026

  February 2026: a developer reported that Claude Code deleted a production SQLite database during a schema migration, after taking it out of Git and turning off backups, leaving "no recovery path."

  Source: GitHub issue #27448 in anthropics/claude-code, February 21, 2026

  A single user's report; the reporter says the data was never recovered.

  Copy citation
• 394 MB

  A 394 MB pg_dump backup, made 30 minutes earlier to cover a rollback window, was deleted unprompted by Claude Code in the middle of a database migration, per a July 2026 bug report.

  Source: GitHub issue #80759 in anthropics/claude-code, “[MODEL] Sonnet spontaneously deleted database backup”, July 24, 2026

  A single user's report; it was closed for inactivity without a reply from Anthropic.

  Copy citation
• 97%

  97% of permission prompts in Claude Code get approved by users, Anthropic says (August 2026).

  Source: Anthropic, “Auto mode is now the default in Claude Code for Pro, Max, and Team plans”, August 7, 2026

  Anthropic's own data about its own product.

  Copy citation
• 17%

  17% of 52 real overeager actions from Anthropic's internal sessions slipped past its auto-mode safety classifier; its incident log includes "attempting migrations against a production database."

  Source: Anthropic, “How we built Claude Code auto mode: a safer way to skip permissions”, March 25, 2026

  Anthropic's own test; the 52 cases were picked from sessions where an employee denied the action or it was flagged afterward.

  Copy citation
• 13.6%

  13.6% of clearly dangerous commands were caught by human reviewers in Anthropic's own test with 1,053 paid testers, compared with 89% for its auto-mode classifier (August 2026).

  Source: Anthropic, “Auto mode is now the default in Claude Code for Pro, Max, and Team plans”, August 7, 2026

  Anthropic's own test (143 of 1,053 caught by people); approved commands were never actually run.

  Copy citation

Code quality and security

• 45%

  45% of AI-generated code samples failed security tests and introduced OWASP Top 10 vulnerabilities, across more than 100 LLMs Veracode tested in 2025.

  Source: Veracode, “2025 GenAI Code Security Report”, July 30, 2025

  Vendor research by Veracode.

  Copy citation
• 86%

  86% of the time, AI tools failed to defend against cross-site scripting in relevant code samples, and Java code had a 72% security failure rate (Veracode, 2025).

  Source: Veracode, “2025 GenAI Code Security Report”, July 30, 2025

  Vendor research by Veracode; these are failure rates, not pass rates.

  Copy citation
• 44%

  44% of AI code generation tasks introduced a risky security vulnerability in Veracode's 2026 tests; the average security pass rate barely moved, from 55% to 56%.

  Source: Veracode, “2026 GenAI Code Security Report: AI Is Writing More of Your Code but Security Hasn't Caught Up”, July 28, 2026

  Vendor research by Veracode.

  Copy citation
• 15%

  15% was the security pass rate on cross-site scripting tasks, and 12% on log injection, in Veracode's 2026 tests.

  Source: Veracode, “2026 GenAI Code Security Report: AI Is Writing More of Your Code but Security Hasn't Caught Up”, July 28, 2026

  Vendor research by Veracode; these are pass rates, not failure rates.

  Copy citation
• 11.8%

  11.8% of SWE-Agent + Claude 4 Sonnet solutions were secure, though 57% worked, on 186 real-world feature tasks where human developers had once shipped a vulnerability (SusVibes benchmark, ICML 2026).

  Source: Zhao et al. (CMU and others), “Is Vibe Coding Safe? Benchmarking Vulnerability of Agent-Generated Code in Real-World Tasks”, September 21, 2026

  Figures are from the current arXiv version (v4) and the ICML paper; the first version used 200 tasks and reported 10.5% secure.

  Copy citation
• Almost half

  Almost half of the code snippets five LLMs wrote contained bugs that could potentially be exploited, using prompts Georgetown CSET designed to surface insecure code (November 2024).

  Source: Georgetown CSET, “Cybersecurity Risks of AI-Generated Code”, November 2024

  Copy citation
• 1.7×

  1.7× more issues showed up in AI-co-authored pull requests than in human-only ones, across 470 open-source PRs (CodeRabbit, December 2025).

  Source: CodeRabbit, “State of AI vs Human Code Generation Report”, December 17, 2025

  Vendor research by CodeRabbit (320 AI-co-authored PRs, 150 human-only), counted with its own issue categories.

  Copy citation
• Up to 2.74×

  Up to 2.74× more security issues turned up in AI-co-authored pull requests than in human-only ones (CodeRabbit, 2025).

  Source: CodeRabbit, “State of AI vs Human Code Generation Report”, December 17, 2025

  Vendor research by CodeRabbit; 2.74× is the upper end, not an average.

  Copy citation
• 96%

  96% of developers don't fully trust that AI-generated code is functionally correct, yet only 48% say they always check it before committing (Sonar, January 2026).

  Source: Sonar, “Sonar Data Reveals Critical 'Verification Gap' in AI Coding: 96% Don't Fully Trust Output, Yet Only 48% Verify It”, January 8, 2026

  Vendor survey by Sonar of more than 1,100 developers.

  Copy citation
• 42%

  42% of committed code is now AI-generated, developers told Sonar, and they expect it to reach 65% by 2027.

  Source: Sonar, “Sonar Data Reveals Critical 'Verification Gap' in AI Coding: 96% Don't Fully Trust Output, Yet Only 48% Verify It”, January 8, 2026

  Vendor survey by Sonar; self-reported, not measured from codebases.

  Copy citation
• 66%

  66% of developers say their biggest frustration with AI is output that's "almost right, but not quite" (Stack Overflow, 2025).

  Source: Stack Overflow, “2025 Developer Survey: AI”, 2025

  Copy citation
• 8.3% to 12.3%

  Copy-pasted lines rose from 8.3% to 12.3% of changed code between 2021 and 2024, while refactoring-related "moved" lines fell from 25% to under 10%, in GitClear's analysis of 211 million changed lines.

  Source: GitClear, “AI Copilot Code Quality: 2025 Data Suggests 4x Growth in Code Clones”, 2025

  Vendor research by GitClear; it shows a correlation, not a cause.

  Copy citation
• 2,000+ vulnerabilities

  2,000+ vulnerabilities, 400+ exposed secrets and 175 instances of exposed personal data turned up when Escape scanned 5,600 publicly available vibe-coded apps (October 2025).

  Source: Escape, “Methodology: How we discovered vulnerabilities in apps built with vibe coding”, October 29, 2025

  Vendor research by Escape; about 4,000 of the apps were built with Lovable.

  Copy citation
• 170 of 1,645

  170 of 1,645 apps showcased on Lovable's "Launched" page (10.3%) had row-level security so weak that strangers could reach user data, across 303 exposed endpoints, according to researcher Matt Palmer, a Replit employee at the time (CVE-2025-48757, May 2025).

  Source: Matt Palmer, “Statement on CVE-2025-48757”, May 29, 2025

  Self-published by the researcher; Replit competes with Lovable.

  Copy citation
• CVSS 9.3

  9.3 out of 10 ("critical") is the CVSS score on CVE-2025-48757, the Lovable row-level security flaw, though Lovable disputes it and says securing data is each customer's job.

  Source: CVE Program, “CVE-2025-48757”, May 30, 2025

  Copy citation
• Under 24 hours

  Under 24 hours is how long Wix took to fix a Base44 flaw that could have let outsiders into private apps using nothing but a non-secret app_id (Wiz, July 2025). Wix found no evidence it was exploited.

  Source: Wiz Research, “Critical Vulnerability in AI Vibe Coding platform Base44”, July 29, 2025

  Vendor research by Wiz.

  Copy citation
• 1.5M tokens

  1.5 million API tokens and 35,000 email addresses were exposed by Moltbook, a vibe-coded social network whose Supabase database had no row-level security; it was fixed within about three hours of Wiz's report (Wiz, February 2026).

  Source: Wiz Research, “Hacking Moltbook: AI Social Network Reveals 1.5M API Keys”, February 2, 2026

  Vendor research by Wiz.

  Copy citation
• 18,697 records

  18,697 user records were exposed by one Lovable-hosted app with 16 security flaws, six of them critical, according to researcher Taimur Khan (February 2026).

  Source: The Register, “Lovable-hosted app littered with basic flaws exposed 18K users, researcher claims”, February 27, 2026

  Lovable's CISO said it acted "within minutes" of the report.

  Copy citation
• 76 days

  76 days (Feb 3 – Apr 20, 2026) is how long chat histories and source code of public Lovable projects could be read by any Lovable user with the link, according to Lovable.

  Source: Lovable, “Our response to the April 2026 incident”, April 22, 2026

  Lovable's own account; it says a fix shipped within two hours of the public report and private projects were never affected.

  Copy citation
• 48 days

  48 days is how long the researcher said their report of the Lovable exposure had sat unfixed before they went public on April 20, 2026.

  Source: @weezerOSINT on X, April 20, 2026

  The researcher's figure; Lovable says the first HackerOne report came in on February 22, which makes it 57 days.

  Copy citation
• One ticket

  One planted support ticket was enough to make a Cursor agent with Supabase service_role access leak a private tokens table, in General Analysis's July 2025 demo.

  Source: General Analysis, “Supabase MCP Security: How Prompt Injection Leaked Private Tables”, July 8, 2025

  Vendor demo using fake data; Supabase says no customer has reported a data leak via MCP.

  Copy citation
• v1.84.0

  1.84.0 is the version of Amazon Q Developer's VS Code extension that shipped with injected malicious code in July 2025; AWS says a syntax error kept it from running.

  Source: AWS, “Security Bulletin AWS-2025-015”, July 23, 2025

  Fixed in version 1.85.0; AWS says no services or customer environments were affected.

  Copy citation

Where the term came from

• Feb 2, 2025

  February 2, 2025: Andrej Karpathy coined "vibe coding" in a post on X, describing a way of building where he hits "Accept All" and stops reading the diffs.

  Source: Andrej Karpathy on X, February 2, 2025

  Copy citation
• 2025

  "Vibe coding" was named Collins Dictionary's Word of the Year for 2025, less than a year after Karpathy coined it.

  Source: PA Media (via The Irish News), “Collins' Word of the Year for 2025 revealed”, November 6, 2025

  Copy citation

Developer adoption

• 84%

  84% of developers were using or planning to use AI tools in 2025, and 51% of professional developers used them every day.

  Source: Stack Overflow, “2025 Developer Survey: AI”, 2025

  Based on 33,662 answers to this question, 26,004 of them from professional developers.

  Copy citation
• 72%

  72% of respondents said vibe coding is not part of their work, and another 5% answered "No, emphatically" (Stack Overflow, 2025).

  Source: Stack Overflow, “2025 Developer Survey: AI”, 2025

  Share of the 26,564 people who answered this question (54.2% of all survey respondents), not only professional developers.

  Copy citation
• 46%

  46% of developers actively distrust the accuracy of AI tools, while 33% trust it (Stack Overflow, 2025).

  Source: Stack Overflow, “2025 Developer Survey: AI”, 2025

  Copy citation
• 73%

  73% of respondents to Stack Overflow's 2026 survey use coding assistants and agents daily.

  Source: Stack Overflow, “2026 Developer Survey: AI”, 2026

  30,903 responses from 169 countries, collected June 23 to August 5, 2026.

  Copy citation
• 90%

  90% of tech professionals use AI at work, but 30% have little or no trust in the code it generates (Google DORA, 2025).

  Source: Google Cloud, “Announcing the 2025 DORA Report”, September 23, 2025

  Survey of nearly 5,000 technology professionals.

  Copy citation

Startups

• 25%

  25% of startups in Y Combinator's Winter 2025 batch had codebases that were 95% AI-generated, according to YC managing partner Jared Friedman.

  Source: TechCrunch, “A quarter of startups in YC's current cohort have codebases that are almost entirely AI-generated”, March 6, 2025

  TechCrunch notes the 95% excludes code from imported libraries.

  Copy citation

Revenue and valuations

• 8 months

  8 months is how long Lovable took to go from $1M to $100M in annual recurring revenue (July 2025).

  Source: Lovable, “$100M ARR & Lovable Agent”, July 23, 2025

  Company-reported; counted from Lovable's first $1M in ARR, not from its launch.

  Copy citation
• $13.3B

  $13.3 billion was Lovable's valuation after a $400M Series C in August 2026; users had built more than 60 million projects on it.

  Source: Lovable, “We just raised $400M in Series C funding to help people run their businesses”, August 12, 2026

  Company-reported; the project count is cumulative since November 2024.

  Copy citation
• $1B

  $1 billion in annualized revenue: Cursor crossed it when it raised a $2.3B Series D at a $29.3B valuation in November 2025.

  Source: Cursor (Anysphere), “Series D”, November 13, 2025

  Company-reported.

  Copy citation
• $9B

  $9 billion: Replit's valuation after a $400M round in March 2026, up from $3 billion in September 2025 (Forbes; Replit).

  Source: Forbes (via Replit), “Meet The $9 Billion AI Company Reimagining Vibe Coding”, March 11, 2026. Also: Replit, September 10, 2025

  Company-reported.

  Copy citation
• $80M

  $80 million (plus earn-outs) is what Wix paid for vibe-coding startup Base44 in June 2025; by May 2026, Base44 was at about $150 million in ARR.

  Source: Wix, “Wix Further Expands into Vibe Coding with Acquisition of Base44, a Hyper-Growth Startup that Simplifies Web and App Creation with AI”, June 18, 2025. Also: Wix, “Wix Reports First Quarter 2026 Results”, May 13, 2026

  Company-reported.

  Copy citation

Productivity studies

• 19% longer

  19% longer: that's how much more time experienced open-source developers took on tasks when allowed to use AI tools, in METR's early-2025 randomized trial.

  Source: METR, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity”, July 10, 2025. Also: METR, “We are Changing our Developer Productivity Experiment Design”, February 24, 2026

  16 developers and 246 issues. METR now labels these results out of date: its February 2026 update estimates about 18% less time per task for the 10 returning developers using late-2025 tools, with a range that still includes no effect.

  Copy citation
• ~18% less time

  About 18% less time per task: METR's February 2026 estimate for the 10 developers who returned from its 2025 study, now using late-2025 AI tools. METR reports it as a speedup of -18% with a range of -38% to +9%, which still includes no effect, and says its data is "only very weak evidence" for the size of the change.

  Source: METR, “We are Changing our Developer Productivity Experiment Design”, February 24, 2026

  The estimate covers only the 10 returning developers (137 tasks); 57 developers and 800+ tasks are the totals for the whole 2026 study. METR's chart plots the result as the change in task time when AI is allowed, with negative values marked as speedup.

  Copy citation

Supabase and AI app builders

• 60%+

  60%+ of new Supabase databases are launched by some kind of AI tool, and database launches grew 600% in a year (Supabase, June 2026).

  Source: Supabase, “Supabase Series F”, June 4, 2026

  Company-reported.

  Copy citation
• 40,000+ a day

  40,000+ new databases launch on Supabase every day, and every Lovable Cloud project runs on Supabase behind the scenes (Supabase, September 2025).

  Source: Supabase, “Lovable Cloud + Supabase: The Default Platform for AI Builders”, September 29, 2025

  Company-reported.

  Copy citation
• Millions

  Millions of Supabase projects were created in just a few months through Supabase for Platforms, by AI builders including Lovable, Bolt.new and Baidu's MeDo (Supabase, December 2025).

  Source: Supabase, “Introducing Supabase for Platforms”, December 5, 2025

  Company-reported; Supabase didn't break the total down by platform.

  Copy citation
• Sep 2025

  "Never connect AI agents directly to production data," Supabase's CISO wrote in September 2025, adding that vibe coders "are developing on production databases" instead of separate staging environments.

  Source: Supabase, “Defense in depth for MCP servers”, September 16, 2025

  Copy citation
• Nearly 10M

  Nearly 10 million developers build on Supabase, which raised $500M at a $10B pre-money valuation in June 2026.

  Source: Supabase, “Supabase Series F”, June 4, 2026

  Company-reported.

  Copy citation

How we built this list

We only kept numbers we could trace to whoever produced them: the survey, the research paper, the company announcement, or the incident write-up from the people involved. If we could only find a number on a statistics roundup site, it isn't here.

Incidents include how they ended. Several stories that are retold as permanent data loss ended with the data restored, and we say so. When a study was run by a vendor that sells into this market, the note under the number says so.

BackupDrill sells backups and restore drills for Supabase projects, so we have an interest in this subject. We kept every number as its source states it.