In December 2025, we surveyed 3,200 professional developers across 47 countries about their AI coding agent usage, security practices, and concerns. The results paint a vivid picture of an industry racing to adopt AI tools while struggling to contain the security consequences.
Methodology
We distributed our survey through developer communities on Reddit, Hacker News, LinkedIn, and our newsletter. Respondents were screened to include only developers who had used AI coding agents at least five times in the preceding month.
- 3,200 qualified responses across 47 countries
- 62% work at companies with 50+ engineers
- Margin of error: ยฑ1.7% at 95% confidence
Finding 1: AI Agent Adoption Is Nearly Universal
Among our respondents, 89% use AI coding agents at least once daily. Compared to Stack Overflow's 2024 survey (61% used AI tools at all), this represents a dramatic shift in 18 months.
| Usage Frequency | % | |
|---|---|---|
| Multiple times daily | 61% | |
| Once daily | 28% | |
| Several times weekly | 8% | |
| Weekly or less | 3% |
Finding 2: Security Anxiety Is High, But Action Is Low
67% expressed concern about credential exposure through AI agents. Yet only 23% have implemented automated scanning. This "anxiety-action gap" is the most important finding.
When asked why they haven't adopted AI-specific security tools:
- "I didn't know tools existed" โ 41%
- "It would slow down my workflow" โ 33%
- "My company hasn't mandated it" โ 19%
- "I trust the AI" โ 7%
"The 7% who trust AI not to generate insecure code is the most concerning number. Every security researcher agreed that AI models hallucinate security-relevant patterns regularly."
Finding 3: Security Incidents Are Common
69% of developers experienced at least one AI-related security incident in the past year.
| Incident Type | Experienced | Caught Before Deploy |
|---|---|---|
| API key/secret exposed | 34% | 71% |
| Insecure dependencies added | 28% | 45% |
| Hardcoded credentials | 22% | 63% |
| SQL injection introduced | 12% | 58% |
| XSS vulnerability | 9% | 52% |
Finding 4: Traditional Tools Don't Cover AI-Specific Risks
Most respondents rely on tools designed for human-written code:
- GitHub Dependabot โ 72% (dependencies only)
- ESLint security rules โ 45% (JS/TS only)
- SonarQube โ 31% (general quality)
- GitGuardian/gitleaks โ 24% (secret scanning only)
- AI-specific tools โ 11%
None except purpose-built tools address the holistic risk profile of AI-generated code โ including prompt injection, MCP tool abuse, and AI-specific patterns that traditional scanners miss.
Recommendations
For Developers
- Install a pre-commit secret scanner today
- Audit your MCP tool configurations
- Review AI code with the same scrutiny as third-party libraries
For Engineering Managers
- Establish an AI code security policy
- Add AI output scanning to your CI/CD pipeline
- Track AI-related security incidents separately
Conclusion
The state of AI agent security in 2026 is defined by a paradox: rapid adoption paired with slow security maturation. Developers know the risks โ 67% are worried โ but the tooling hasn't caught up. We built ClaudeArmor to close precisely this gap.