Safety & Security

83% of Firms Run Unapproved AI Coding Agents, BlueFlag Research Finds

BlueFlag Security ships six AI agent governance functions as research finds 83% of organisations run unapproved AI coding agents, some with admin rights.

BlueFlag adds AI agent governance tools for developers
BlueFlag adds AI agent governance tools for developersjamesmarkosborne / pixabay
By Elena Vasquez6 min read

Why it matters

  • BlueFlag research found 83% of organisations run AI coding agents that were not formally approved.
  • Nearly 70% of AI coding agents hold administrator permissions; some firms use up to 36 distinct agents.
  • Six new governance functions are generally available, covering agent classification, review checks, AI app cataloguing, spending analysis, shadow AI detection and remediation.
  • Among early adopters, AI writes up to 72% of all commits; in over one in five organisations an agent has merged code without human review.
  • CEO Raj Mallempati said the platform will begin interrupting attack paths as they form starting early next year.

Eighty-three percent of organisations run AI coding agents that nobody formally approved. That figure comes from new BlueFlag Security research drawn from tens of thousands of human, non-human and agentic AI identities across major enterprises — and it frames why the company just shipped six new AI agent governance functions for its software development lifecycle platform.

The stakes are straightforward. AI agents now write, test, review and deploy code inside corporate environments, often with administrator-level permissions and without human oversight. BlueFlag's research found that in nearly 70% of cases, those agents hold administrator permissions. In more than one in five organisations, an autonomous agent has already merged code with no human in the loop.

What does the new platform actually do?

BlueFlag's expansion centers on a problem most security teams can't currently see: which AI agents exist in a development environment, what they touch, and who is accountable for them. The six new functions are generally available now and work through APIs connected to customers' existing tools.

The capabilities break down as follows:

  • Agent classification and approval. The platform discovers AI agents through code changes and build activity, then classifies them by role — whether they write, test, review or deploy code. It also shows whether an agent operates under a developer's identity or as a separate identity, along with the repositories, pipelines, commits and pull requests linked to that agent.
  • Human review checks. The system flags pull requests approved by an AI agent, commits written entirely by an agent, and pipeline runs deployed without human review. Customers can also require human review at the pull request stage before merge or deployment.
  • Visibility into AI-linked applications. A cataloguing function identifies software applications in the codebase that use AI, including which model providers they call, which agent frameworks they use, and whether models run on company infrastructure or through third parties. It measures those applications against the OWASP Agentic Top 10.
  • Spending analysis. Customers can match token consumption and spending to commits, pull requests and accepted lines of code, broken down by developer, tool and model.
  • Shadow AI detection. The platform integrates with endpoint detection and response tools including CrowdStrike and SentinelOne to identify AI tools running on developer devices — without requiring additional software on the endpoint.
  • Automated remediation. Two functions called AccessGuard and DriftGuard let customers address risks through manual instructions, a generated remediation script, or a one-click automated fix that can be rolled back.

The platform remains read-only by default. Automated fixes run only when customers opt in, a design choice that matters for security teams wary of giving a governance tool write access to production pipelines.

Why identity hygiene is the sticking point

AccessGuard and DriftGuard target the misconfigurations that turn AI agents into attack paths. AccessGuard focuses on identity hygiene issues: excessive permissions, stale credentials and dormant team memberships. DriftGuard goes after repository and pipeline misconfigurations, including missing branch protections and settings changes over time.

The distinction maps directly onto BlueFlag's research findings. Fifty-six percent of organisations have autonomous agents operating under their own identity. Some organisations use as many as 36 distinct AI coding agents. Each of those identities represents a credential, a permission set and a potential drift problem that traditional identity governance wasn't built to handle.

Among early adopters, AI writes up to 72% of all commits, according to the company's data. That volume means manual oversight of AI-produced code is already impractical at many organisations, which is precisely the gap the new review controls aim to close.

What did BlueFlag's CEO say?

Raj Mallempati, Chief Executive Officer of BlueFlag Security, framed AI agents as a governance issue that development teams have largely ignored.

"An AI agent is now the most powerful and least governed identity in the development environment," Mallempati said. "That level of trust is at best misguided and at worst a potentially catastrophic threat to the business. BlueFlag offers the unparalleled ability to govern agents where they work, alongside the human developers and non-human identities in the same ecosystem, and connect their behavior to everything around them, so an attack path shows up the moment one manifests."

He also laid out where the product goes next. "Agents are about to become the primary developers of software. People will set the goals, the guardrails, and where a human has to sign off. BlueFlag is already the governance layer across both, and early next year it will begin to interrupt an attack path as it forms, not only report it," Mallempati said.

That last commitment is notable. Moving from detection to active interruption of attack paths would put the platform in a more interventionist role than read-only monitoring, and it puts a rough timeline — early next year — on the shift.

How did a customer react?

Marelli, the automotive parts supplier, is among the companies BlueFlag cited as a platform user. Kristie Pfosi, Global Chief Information Security Officer at Marelli, said the software surfaced AI-related activity in the company's repositories within days and tied it to approval and review gaps.

"We accept some unknowns when the attack comes from outside. We should not have to accept them in our own development environment, and agentic AI has made that gap bigger overnight," Pfosi said.

She described the first-week findings in concrete terms. "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. Those are findings we can act on the same day, and no other tool we run could have surfaced them," Pfosi said.

Why the spending controls matter as much as the security ones

The token-to-output matching function addresses a second, quieter problem: AI development spending that outruns results. By matching token consumption and spend to commits, pull requests and accepted lines of code, the tool is designed to show how much of the codebase AI actually wrote, whether paid seats sit idle, and where spending climbs without corresponding output.

That positions BlueFlag not just as a security vendor but as a cost-accounting layer for AI-assisted development — a capability that matters as enterprises struggle to justify ballooning AI budgets. With AI writing up to 72% of commits among early adopters, knowing which tools produce accepted code versus wasted tokens becomes a procurement question as much as a security one.

The broader context is a market scrambling to govern what it deployed faster than it could control. Shadow AI on developer devices, unapproved agents with admin rights, autonomous merges without review — each finding in BlueFlag's research describes an environment where adoption outran policy. BlueFlag is betting that governance sold where the agents work, inside the development lifecycle itself, wins over retrofitted security tooling, and its roadmap now points toward interception rather than observation.

Source: itbrief.com.au

Share this article:

More from Elena Vasquez

Elena Vasquez

Show full bio

Market editor covering technology and product news in artificial intelligence and machine learning at AI In Context.

708 articles

Related articles

  1. Okta-Led Blueprint Alliance Wants a Kill Switch for Every AI Agent
  2. 87% of Enterprise Developers Use Low-Code. That Is the Agent Security Problem
  3. Australian firm opens the books on five months of AI agent controls
  4. PwC survey: AI accountability scattered, 11% say it's unclear
  5. Uptiq launches GovernAI to govern AI agents in banking at no extra cost

« Previous articleNext article »