What's new in AI app security.
A short, sourced roundup. Last updated 15 October 2026. Every figure links to its source.
AI-service secret leaks up 81%
GitGuardian’s State of Secrets Sprawl 2026 (March 2026) counted 28.65M new hardcoded secrets in public GitHub commits in 2025 (+34% YoY) and 1,275,105 leaked AI-service secrets (+81%). It also found 24,008 unique secrets in MCP configuration files, and that 64% of valid secrets from 2022 were still not revoked.
Source: GitGuardian ↗AI-assisted commits leak about 2× more
The same report found commits co-authored by Claude Code leaked secrets at ~3.2% vs a 1.5% baseline — while stressing developers still decide what gets pushed.
Source: GitGuardian ↗Thousands of vibe-coded apps exposed data
RedAccess reported to Axios and WIRED that it found ~380,000 public assets built with Lovable, Base44, Replit and Netlify, about 5,000 with virtually no authentication and sensitive corporate data, including medical and financial records.
Source: Axios, May 2026 ↗ · WIRED ↗Open databases beat leaked keys
A scan of 30,998 live vibe-coded apps (Aug 2026) found notable keys in 1,332 apps (mostly Google keys), OpenAI keys in 33, Supabase service_role in 3 — but 2,096 had a Supabase table readable with no login.
Source: Reeve ↗ChatGPT keys in live websites
Cyble Research found 5,000+ public GitHub repos and ~3,000 live websites leaking OpenAI API keys in source and client-side JavaScript.
Source: Cyble ↗AI coding agents as an attack surface
Check Point Research disclosed CVE-2025-59536 and CVE-2026-21852: malicious Claude Code project files (hooks, MCP, ANTHROPIC_BASE_URL) could run commands or exfiltrate API keys. Anthropic patched both before disclosure (Feb 2026).
Source: Check Point Research ↗