GitGuardian Report: AI-Assisted Coding Led to 28.65M Leaked Secrets in 2025, 34% Year-Over-Year Spike
Key Takeaways
- ▸AI-assisted coding accelerated software development by 43% in 2025 but created proportionally larger security vulnerabilities, with 28.65M new hardcoded secrets leaked—the largest single-year increase on record
- ▸AI service secrets represent the fastest-growing category of leaks with an 81% year-over-year increase, driven by rapid adoption of new AI providers, wrappers, and integrations without mature security practices
- ▸Documentation and quickstart guides that normalize hardcoded credentials in MCP and other AI infrastructure are accelerating ecosystem-wide secret sprawl before security teams can establish controls
Summary
GitGuardian's latest "State of Secrets Sprawl" report reveals that 28.65 million hardcoded secrets were added to public GitHub commits in 2025, representing a 34% increase year-over-year and the largest single-year jump recorded. The surge coincides with explosive growth in AI-assisted coding adoption, which made software development faster and more accessible but also expanded the attack surface. Public GitHub commits climbed 43% to 1.94 billion, while the developer base grew 33%, accelerating the pace at which new tools, APIs, and service accounts were deployed without adequate governance.
The report identifies AI service secrets as a particular vulnerability, with AI service credentials leaking at an 81% year-over-year increase and reaching 1.275 million exposures in 2025. Eight of the ten fastest-growing secret detectors were tied to AI services, and LLM infrastructure including orchestration, RAG, and vector storage leaked 5× faster than core model providers. Claude Code-assisted commits showed a 3.2% secret-leak rate versus a 1.5% baseline, though the report emphasizes that developers retain control over accepting or rejecting AI-generated code.
A critical finding involves Model Context Protocol (MCP) configuration files, where 24,008 unique secrets were exposed across public GitHub, including 2,117 valid credentials. The problem is compounded by documentation that encourages unsafe practices, with popular MCP setup guides often recommending hardcoded API keys in configuration files. GitGuardian also notes that internal repositories remain a much larger reservoir of secrets sprawl, with internal repos approximately 6× more likely than public ones to contain hardcoded secrets, creating long-term security debt.
- Internal repositories pose a greater risk than public ones, with 6× higher likelihood of containing secrets, creating a long-term security debt that becomes exploitable once internal systems are compromised
Editorial Opinion
The 2025 secrets sprawl explosion reveals a fundamental tension in AI-driven development: tools that democratize coding and accelerate shipping inevitably outpace security governance. While the 3.2% leak rate for Claude-assisted commits is marginally higher than baseline, this misses the larger systemic issue—the ecosystem is building critical infrastructure on foundations of convenience-first examples. As new AI standards and tools arrive at unprecedented speed, embedding secure-by-default practices from day one is not optional; it's the only way to prevent a generation of vulnerabilities from becoming the legacy of this transformative period.
