BlueFlag AI has published a developer risk study drawn from pooled organizational data, and its headline result concerns how long repository tokens stay alive. Among tokens examined, 99% carried no expiration date. Four in five permitted code changes, and 47% sat unused over ninety days without revocation.
Standing access that nobody revokes
The report blames ordinary churn: permissions pile up as engineers join projects, switch teams or take on admin duties, and contract grants often survive delivery. On average a developer held 61% more access than the job needed. The heaviest unused allocations sat with people who could already reach business-critical repositories, roughly 88% of that group, spanning production services, payment systems and regulated workloads. Developer identities dormant over ninety days yet holding admin rights appeared at four in five organizations.
Machines outnumber the people
Count bots, service accounts, apps and tokens as identities and the ratio turns stark: non-human accounts able to reach code outnumbered human developers eight to one. Machine-held permissions reached critical repositories too, amplifying what a stolen secret is worth.
Agentic coding adds volume. The median organization used about seven AI coding tools, a few ran more than 30, and 83% of those measured had unapproved agents somewhere. Most kept no approved-tool list. Assistants ride a developer identity and take its rights along; autonomous agents stand alone and write, review and merge without signoff.
Of the identities agents ran as, 68% held admin privileges and 81% skipped behavioral monitoring. In most organizations running agents, six of ten, those agents wielded tokens with clone rights over every repository.
Because agent actions surface under a human’s account name, commit tags and metadata are the only way to tell who did what. Fixes are simple: expiry by default, deleting idle service accounts, owning each agent’s scope.
