AI security statistics, with a source for every one.

new secret keys pushed to public GitHub code in 2025

That was up 34% in one year, the biggest jump GitGuardian has recorded.

Source: GitGuardian, State of Secrets Sprawl 2026 · March 2026

rise in leaked keys for AI services in a year

About 1.28 million AI-service keys turned up in public code in 2025.

Source: GitGuardian, State of Secrets Sprawl 2026 · March 2026

still-working keys found in public AI-agent (MCP) setup files

They came from 24,008 secrets in MCP config files, partly because setup guides tell people to paste keys straight in.

Source: GitGuardian, State of Secrets Sprawl 2026 · March 2026

of keys leaked back in 2022 still worked in January 2026

Deleting a key from your code does nothing. Only revoking it does.

Source: GitGuardian, State of Secrets Sprawl 2026 · March 2026

place for prompt injection on the OWASP Top 10 risks for AI apps

OWASP is the non-profit whose lists security teams actually use to decide what to fix first.

Source: OWASP GenAI Security Project, LLM01:2025 Prompt Injection · 2025

successful attacks on AI agents in one month-long public contest

Red teamers sent 1.8 million prompt-injection attempts at agents built on 22 top AI models, and every agent could be broken.

Source: NeurIPS 2025, Security Challenges in AI Agent Deployment · December 2025

of hidden-instruction attacks worked on a browser AI agent with no extra defenses

Anthropic tested 123 attack cases before launching Claude in Chrome and added defenses because of it.

Source: Anthropic, Piloting Claude in Chrome · August 2025

of 2,614 MCP servers use file code prone to “path traversal” bugs

Path traversal is when a tool meant to read one folder can be tricked into reading any file on your computer.

Source: Endor Labs, Classic Vulnerabilities Meet AI Infrastructure · January 2026

of 2,614 MCP servers use code linked to command-injection bugs

Command injection means a bad input can make the tool run commands on your machine.

Source: Endor Labs, Classic Vulnerabilities Meet AI Infrastructure · January 2026

of people clicked fully AI-written phishing emails in a study

A small study (101 people), but the gap is big: that matched human experts, and was more than four times the 12% who clicked generic phishing emails.

Source: Heiding, Schneier et al., Evaluating LLMs’ Ability to Launch Automated Spear Phishing (arXiv) · November 2024

lost to cybercrime reported to the FBI in 2024

A record, up 33% from 2023, across 859,532 complaints.

Source: FBI Internet Crime Complaint Center, 2024 IC3 Report · April 2025

lost to business email compromise in 2024

That is the fake “boss” or “vendor” email asking you to wire money, and it was the second-costliest crime type.

Source: FBI Internet Crime Complaint Center, 2024 IC3 Report · April 2025

phishing and spoofing complaints to the FBI in 2024

The most-reported crime type of the year.

Source: FBI Internet Crime Complaint Center, 2024 IC3 Report · April 2025

of the top 1.8 million domains have a DMARC record

So roughly half of popular domains still have none of the main protection against someone faking their email.

Source: EasyDMARC, 2025 DMARC Adoption Report · 2025

higher success rate for fake texts and calls than for email phishing

Source: Verizon, 2026 Data Breach Investigations Report · May 2026

of employees now use AI tools often, up from 15% a year earlier

Source: Verizon, 2026 Data Breach Investigations Report · May 2026

of people who use AI at work bring their own tools

Tools your company never approved are tools nobody is checking.

Source: Microsoft & LinkedIn, 2024 Work Trend Index · May 2024

of companies hit by an AI-related breach lacked proper AI access controls

Source: IBM, Cost of a Data Breach Report 2025 · July 2025

extra breach cost for companies with lots of unapproved “shadow AI”

Source: IBM, Cost of a Data Breach Report 2025 · July 2025

of breached companies had no AI governance policy

In other words, no written rules about which AI tools staff may use or what data they can paste in.

Source: IBM, Cost of a Data Breach Report 2025 · July 2025