Software development security startup BlueFlag Security Inc. today expanded the artificial intelligence agent governance tools in its developer risk platform to cover unapproved coding agents, including those running on engineers’ laptops outside the code repository.

The update also brings automated fixes for some of the risks the platform turns up.

BlueFlag is going after a problem that has grown fast as coding assistants have evolved into agents that plan their own work and spin up the environments they need. Many sign in with a developer’s credentials. An agent working that way inherits every permission the developer holds.

Raj Mallempati, BlueFlag’s co-founder and chief executive, described the AI agent as “the most powerful and least governed identity in the development environment.” The company plans to start interrupting attack paths as they form early next year, he said. For now, the platform reports them.

In the current release, the work starts with finding the agents, which BlueFlag does from activity in the code. It picks up each agent through the commits and pull requests it produces, then classifies it by what it does with the code and whose identity it uses. Security teams can trace any agent back to the developers who work with it and the repositories it operates in, with unapproved agents marked as such.

AI tools running locally on a developer’s machine are harder to spot. To catch those, BlueFlag reads data from endpoint detection and response products such as those from CrowdStrike Holdings Inc. and SentinelOne Inc., so nothing new has to be installed on the device.

Human review gets its own checks. BlueFlag now flags pull requests approved by an AI agent, along with commits an agent wrote from start to finish and pipeline runs that went out with no person signing off. Organizations can also require a human review before anything merges or deploys. For any software release, customers can document where a person was in the decision and where one was not.

A separate inventory looks at AI inside the applications a company builds. It records the model providers and agent frameworks each application relies on and scores them against the Open Worldwide Application Security Project’s Top 10 for Agentic Applications.

AI spending is measured as well. BlueFlag ties the tokens and dollars each developer burns through to the commits and accepted code that came out of them, and it points to places where spending keeps climbing with no new commits behind it. Paid seats that sit idle show up in the same report.

Remediation runs through two new tools. AccessGuard targets identity problems such as overpermissioned accounts and stale credentials.

Misconfigured repositories and pipelines are handled by DriftGuard, which corrects settings such as missing branch protections that drift over time. Fixes can be applied by hand from written instructions or run automatically with one click, and automated changes can be rolled back.

Auto parts supplier Marelli Holdings Co. Ltd. is a customer. Kristie Pfosi, its global chief information security officer, said agentic AI made the unknowns inside her company’s own development environment bigger “overnight.”

“Within the first week BlueFlag showed us which AI agents were working in our repositories, which developers they worked with, and where the gaps were in how they were approved, scoped and reviewed,” Pfosi said. “Those are findings we can act on the same day.”

The launch comes a day after BlueFlag published its first research report, “The State of Developer Risk 2026.” Its figures come from metadata the platform collected on tens of thousands of identities and access tokens at a subset of BlueFlag’s enterprise customers, in a snapshot taken in September.

Unapproved coding agents turned up in 83% of the organizations measured, and more than two-thirds of the identities agents operate under carry administrator rights. In 22% of organizations an autonomous agent had already merged code with no human involved.

Older problems have not gone away. Nonhuman identities that can reach code repositories outnumber human developers eight to one, and 99% of access tokens are set never to expire.

The new features are generally available now. BlueFlag says the platform connects through application programming interfaces that are read-only by default and can be running in under five minutes, with automated fixes applied only when a customer opts in.

The company launched in March 2024 with $11.5 million in seed funding. Maverick Ventures and Ten Eleven Ventures led that round and came back to lead a Series A this past March that took total funding to $28 million.

Image: BlueFlag