Our first research report, The State of Developer Risk 2026, measures the whole environment that builds software: the identities acting on the code, the tools they work through, and the pipelines that carry it to production. Here is what we found.
Security teams have spent years finding vulnerabilities in code: static analysis, dynamic analysis, software composition analysis, a whole industry built around the line of code.
Attackers noticed, and they went around it.
The breaches that defined SDLC security this year did not begin with a vulnerable line. They began with an identity: a stolen CI token, a service account used to tamper with build pipelines, a developer’s device giving up the tokens that cloned thousands of repositories, an AI coding agent turned into a credential thief. In BlueFlag’s experience across customer engagements, 90% of SDLC attacks start with the identities acting on the code and the tools they interact with, not in the code itself.
So for our first research report we did not ask security leaders what they think is happening inside their development environment. We measured it, from production data across a representative subset of large enterprises, aggregated and anonymized. 25 findings in 6 chapters. Here is a sample of what the data says.
1. Human Developers: Permissions Few Take Back
Permissions get added with every project and team change, and rarely taken away.
- 61% of the permissions granted to the average developer are never used.
- 88% of the most over-provisioned developers can reach the organization’s business-critical repositories.
- 4 in 5 organizations have dormant developer identities that still hold administrator permissions.
2. Non-Human Identities: The 8:1 Majority Almost No One Reviews
Service accounts, access tokens, bots, and connected applications keep the pipelines running, and they get the least attention.
- For every human developer who can reach the organization’s repositories, 8 non-human identities can too.
- 99% of access tokens have no expiration date, and 84% can write code, not only read it.
- 47% of access tokens have sat unused for more than 90 days and have never been revoked.
3. AI Coding Agents: Superhuman Reach, Almost No One Watching
They are the most powerful identities in the environment and fall into two categories: coding assistants, which work under a developer’s identity and inherit its access and permissions, and autonomous agents, which operate as their own identity and can write, review, and merge code with no human in the loop.
- Up to 72% of all commits among the early adopters are now written by AI.
- 83% of organizations run AI coding agents (GitHub Copilot, Claude, and Cursor among them) that were never formally approved.
- 68% of the identities AI coding agents operate under hold administrator permissions.
- 81% of those identities have no one monitoring what they do.
- 56% of organizations have autonomous AI agents operating as their own identity.
4. Toolchain Controls: Switched Off, Bypassed, and Code Walking Out the Door
The settings that keep risky code out of production are frequently missing or bypassed, and when code is taken, no one sees it leave.
- 4 in 10 repositories leave their main branch unprotected, so any change can ship with no review.
- 4 in 5 organizations let administrators bypass code review, thousands of times in a typical affected organization.
- 2 in 3 organizations merged code with reviewer comments still open.
- In 50% of organizations, an identity suddenly cloned far more code than it ever had before, which is what code theft can look like from the inside.
Alone, a Finding. Together, an Attack Path.
Each of these findings warrants attention on its own. Their full significance appears when they connect. In sequence across an identity, a tool, and a pipeline, they can form an attack path: source code leaving with a departing contractor, a stolen credential operating from another continent, or an autonomous agent shipping code no human reviewed. The signals live in different systems, and in most organizations nothing connects them.
Nearly every finding in the report has a known remedy. The hard part is not fixing the risk. It is seeing it.
Find Out Who, and What, Is Acting on Your Code
The full report holds all 25 findings and sets the baseline. A new edition follows every 6 months, with more organizations and more data on AI coding agents, so the series shows not only where they stand today but whether they are getting better.
