Skip to content

Vibe-coding vulnerability surge

Signal card · maintained from public sources · last updated 2026-09-10

CategoryTechnology & AI
StatusStrengthening
Confidence🟩
First observed2026-04-04
Last updated2026-09-10

Overview

Security researchers are documenting a quantified, accelerating surge in vulnerabilities traceable directly to AI-generated ("vibe coded") software — not just anecdotal concern about AI code quality, but tracked CVE counts and platform-level vulnerability audits. Relevant to any portfolio company whose engineering org has adopted AI coding assistants without a matching increase in review/security capacity — the accumulating security debt is a real, measurable liability, not a hypothetical one.

Evidence log

  • 2026-04-04 — Cloud Security Alliance research note: 45% of AI-generated code samples introduce OWASP Top 10 vulnerabilities across 100+ tested models with no improvement despite vendor claims; AI-assisted developers commit 3-4x faster but introduce security findings at 10x the rate. Georgia Tech's "Vibe Security Radar" tracked CVEs directly attributable to AI coding tools escalating from 6 in January 2026 to 35 by March 2026, with researchers estimating the true unconfirmed count 5-10x higher across open-source repos. A related audit of 1,400 applications built on vibe-coding platforms found 2,038 critical vulnerabilities, 400+ leaked secrets, and 175 instances of exposed sensitive data (including medical records and credentials). Separately, ~20% of AI-generated code samples reference non-existent packages, enabling "slopsquatting" supply-chain attacks. Cloud Security Alliance
  • 2026 (ongoing) — Trade press covering the same risk category from multiple angles, including a warning that vibe coding could cause "catastrophic 'explosions'" of security failures during 2026 as adoption outpaces review discipline. The New Stack
  • 2026-03 — Veracode's March 2026 "State of Software Security" update measured a 2.74x higher vulnerability rate in AI-generated code versus human-written code, with the overall security pass rate holding flat at approximately 55% — no improvement despite increased vendor quality claims. This is the second independent quantitative dataset (after CSA/Georgia Tech April data), from a named enterprise security firm with its own application testing pipeline. Veracode / BeyondScale Enterprise Guide 2026
  • 2026-07-31 — First named breach incidents attributable to AI-generated code reported by vibe-coding security research aggregators: (1) Tea App Breach — consumer app users' private messages exposed to strangers via AI-generated code that shipped without security review (Barracuda Networks blog confirmed this as a vibe-coding cautionary case, Dec 2025); (2) Moltbook Incident — 1.5 million authentication tokens and 35,000 email addresses leaked because AI-generated API endpoints returned sensitive data without authorization checks. These are the first named incident reports, not statistical projections — the threshold the April estimate ("roughly even chance it produces a distinct, named incident class within 1-2 quarters") has now been reached. 🟩 Primary press confirmed: Wiz published the Moltbook breach report on Feb 2, 2026, documenting the root cause (Supabase Row Level Security never enabled, anon key embedded in client-side JS, founder publicly stated he "didn't write one line of code"), independently corroborated by Techzine, AI CERTs, and TechPlanet. Confidence upgraded to full 🟩 High — the primary-press-confirmation caveat is resolved. Museum of Vibe Coding Security Research · Wiz Blog — Moltbook · Barracuda — Tea App
  • 2026-08-16 — GitGuardian State of Secrets Sprawl 2026: 28.65 million new hardcoded secrets found in public GitHub commits during 2025, a 34% year-over-year increase. AI-assisted commits show a 3.2% secret-leak rate vs. a 1.5% baseline for human-only commits — AI roughly doubles credential exposure rates. This is the first named large-dataset measurement tying AI commit behavior to secrets sprawl at a quantified scale, from a firm whose entire product category is secrets detection. GitGuardian
  • 2026-08-29 — Updated quantification from two independent sources: OX Security (August 2026) reports 62% of AI-built applications ship with critical security vulnerabilities across a multi-thousand app audit. Georgia Tech's School of Cybersecurity and Privacy independently tested 1,689 AI-generated programs and found 40% contain exploitable vulnerabilities mapping to CWE Top 25, with the AI-assisted commit secret-leak rate confirmed at 3.2% vs. 1.5% baseline (double the human-only rate). CVE counts attributed to AI-generated code have grown from 6 (January 2026) to 15 (February) to 35 (March), with 74 catalogued as traceable to AI coding tools as of March 2026; the true unconfirmed count is estimated at 5–10x higher. The two datasets (OX Security 62% critical vuln rate; Georgia Tech 40% exploitable) bracket the range from the April 2026 CSA/Georgia Tech 45% figure, confirming the trend is not improving. OX Security · Georgia Tech/InfoSecurity Magazine · CSA Research Note Aug 2026
  • 2026-08-16 — Demographic shift — vibe coding is now primarily a non-developer practice: TechTimes (May 2026) reports 63% of vibe-coding users have no coding background. This materially changes the risk posture: the population most at risk of shipping insecure AI-generated code has shifted from developers who can partially review output to non-technical founders and operators who cannot evaluate security implications at all. Escape.tech's audit of 1,400 vibe-coded production apps found 65% with security issues and 58% with at least one critical vulnerability. TechTimes · OX Security

Assessment

Confidence: 🟩 High. Four independent data streams: (1) CSA/Georgia Tech April 2026 — 45% of AI-generated code introduces OWASP Top 10 vulns, CVE count escalating from 6 (Jan) to 35 (Mar); (2) Veracode March 2026 — 2.74x higher vulnerability rate vs. human-written code, 55% pass rate flat; (3) named breach incidents: Moltbook (primary-press confirmed by Wiz, Feb 2, 2026; Supabase RLS never enabled, 1.5M tokens + 35K emails) and Tea App (Barracuda, Dec 2025; 72K government IDs exposed); (4) GitGuardian Secrets Sprawl 2026 — AI commits produce secrets at 2x the baseline rate (3.2% vs. 1.5%). Primary-press-confirmation caveat from July 31 review is now fully resolved. Confidence: unanimous 🟩 High.

Estimate: The named-incident threshold has materialized and is confirmed by primary press. H2 2026 risk: the demographic shift to non-developer vibe coders (63% have no coding background, TechTimes May 2026) materially increases the probability of a major-scale breach reaching trade press as an AI-code-quality story — because non-developers cannot evaluate output security before shipping. When that breach occurs, it will likely force vendor-level AI code review gate requirements and enterprise SLA changes.

Review note (2026-07-31): Veracode March 2026 data (2.74x rate, 55% flat pass rate) provides the second independent quantitative dataset. Tea App Breach and Moltbook Incident are the first named incidents where AI-generated code was the direct causal path to a breach — critical threshold reached, though primary press confirmation pending. Status upgraded Watching → Strengthening; confidence upgraded 🟨 Moderate → 🟩 High.

Review note (2026-08-08): No new named vibe-coding breach incidents found in the Aug 7-8 window. DEF CON 34 AI Village Day 2 (Aug 7) is running research on AI agent security failures via prompt injection in telemetry pipelines, Slack, and phone calls — this tracks adjacent territory (AI-as-attacker) rather than AI-generated code introducing bugs (this signal's scope). No confidence or status change; evidence base unchanged from July 31 review.

Review note (2026-08-29): Two additional independent quantification datasets added: OX Security (62% of AI-built apps ship with critical vulns) and updated Georgia Tech data (40% of 1,689 AI-generated programs exploitable, CWE Top 25). CVE trajectory from 6 (Jan) → 15 (Feb) → 35 (Mar) reconfirmed. Secret-leak rate (3.2% vs. 1.5% baseline) consistent across GitGuardian data. The range of "40–62% of AI-generated code has critical vulnerabilities" across four independent sources (CSA 45%, Veracode 2.74x rate, OX Security 62%, Georgia Tech 40%) is now stable and consistently alarming. No change to Strengthening status or 🟩 High confidence. Next review: Q3 2026 Veracode/CSA dataset update; any named major breach where AI-generated code is the specifically attributed root cause.

Review note (2026-08-16): Moltbook primary press confirmation resolved (Wiz blog, Feb 2, 2026; Supabase RLS root cause confirmed). GitGuardian Secrets Sprawl 2026 adds a fourth independent data stream from a named secrets-detection firm. TechTimes demographic data (63% non-developer users) is a new leading indicator for H2 2026 risk — non-developers who can't review AI output are now the primary vibe-coding population. Three new evidence entries added. No change to Strengthening status or 🟩 High confidence — already at ceiling on both. Next review: monitor for Q3 2026 Veracode/CSA vulnerability-rate dataset update and for any major named breach reaching trade press specifically as an AI-code-quality story.

Graduation review (2026-09-02, SIG-12): verdict — stays active, despite ≥30 days at 🟩 High + Strengthening. Moltbook now has primary press confirmation (Wiz, per the Aug 29 review), closing the earlier sourcing caveat — but the watched event is still ahead: a large-scale named breach reaching the trade press as an AI-code-quality story rather than a generic data breach, which the card's own estimate still frames as the open question. Until that threshold is crossed, the quantified surge lives in research datasets and security-vendor blogs, not mainstream coverage, and the card is still doing early-warning work. Next verdict due at the first signal-review after 2026-10-02.

Review note (2026-09-10): No new named vibe-coding breach or fresh quantitative dataset found in this window. No status or confidence change; evidence base unchanged from the Aug 29 review.

Sources

← All signals