- BlueFlag Security says 99% of the repository access tokens it measured have no expiration date, and 84% can change code.
- The report says AI coding agents often run under powerful identities: 68% of those identities hold administrator permissions.
- Leaders should ask who owns each token and agent, which ones are unused, and whether anyone can tell AI changes from human ones.
Keys to your code may outlive the work they were for
BlueFlag argues the numbers point to a blind spot: many organizations may lack a clear view of which keys are still live. The question sits behind The State of Developer Risk 2026, the first research report from BlueFlag Security.
BlueFlag says it used production data from a representative subset of large enterprises. The data is aggregated and anonymized. The report has 25 findings in six chapters. This story covers only the headline findings.
The idea: the risk has moved from the code to the keys
Security teams spent years hunting flaws in lines of code. BlueFlag says attackers went around that work. Across its customer engagements, it puts the share of attacks on software development that begin with a trusted identity at 90%. The flaw is not in the code. It is in who, or what, is allowed to touch it.
That 90% is BlueFlag's customer experience. It is not a result from the measured sample.
Think of office keys. Hiring hands them out. Few routines collect them. A contractor finishes, a project ends, and the key stays in a drawer. The report points to the same pattern in code repositories, at machine scale.
What the numbers show about tokens and people
An access token is a long secret string. It lets a tool read or change code without a person typing a password. A token with no expiry date works until someone revokes it.
Of the tokens BlueFlag measured, 99% had no expiration date. Some 84% allowed code changes, not just reading. And 47% had gone unused for over 90 days without being revoked.
Tokens are part of a bigger crowd. BlueFlag counts eight non-human identities for each human developer with repository access. Help Net Security notes that this count treats tokens as identities. The rest are service accounts, bots and connected applications.
People also hold access they do not use. BlueFlag says 61% of the permissions given to an average developer sat idle. Among developers with the most unused permissions, 88% could reach business-critical repositories. In four in five organizations, inactive developer accounts still held administrator rights.
AI coding agents work inside the same loose access
BlueFlag describes two kinds of AI coding tool. Assistants borrow a developer's identity and its permissions. Autonomous agents act as identities of their own. They can author, check and merge changes with no human sign-off along the way.
BlueFlag says 83% of organizations run AI coding agents nobody formally approved. GitHub Copilot, Claude and Cursor are among them. Help Net Security adds that a typical organization used about seven such tools. Some used over 30.
Administrator rights sat on 68% of the identities these agents ran under, BlueFlag says. Of those identities, 81% had no monitoring of what they did. Help Net Security also relays a finding on reach. Among organizations with AI activity, agents in six in ten held tokens broad enough to clone every repository.
Credit is another problem. Help Net Security notes that work done through a developer's account can look like that person's own. Without commit tags or similar metadata, human and AI work blur together. BlueFlag says up to 72% of commits among early adopters are now written by AI.
Review checks are often missing or skipped
BlueFlag says four in ten repositories leave the main branch unprotected. Any change can then ship without review. In four in five organizations, administrators were able to skip code review. For a typical affected organization, that happened thousands of times.
In two in three organizations, code was merged while reviewers' comments were still open.
Half the organizations had an identity that suddenly copied far more code than before. In 28%, one identity copied at least 100 repositories it had never touched. Help Net Security stresses that such patterns may merit a closer look. They do not prove misuse or theft.
Keep the limits in mind. This is a vendor report from a company that sells governance tooling. The sample is a subset of large enterprises, and the public summary does not give a full method. Read the figures as BlueFlag's measurement, not an industry average.
Questions to put to your team
First, ask for a list of every token that can change code. Each should have an owner and an expiry date. Second, ask how many have sat unused for 90 days, and who will revoke them.
Third, ask which AI coding tools are in use and which are approved. Ask what permissions each one holds. Fourth, ask whether the team can tell an AI-written change from a human one. Fifth, ask how often administrators skipped code review last quarter.
BlueFlag says nearly every finding has a known remedy. The hard part, it says, is seeing the risk. A key you cannot count is a key you cannot take back.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: BlueFlag Security.
