BlueFlag Security reports excess developer and AI coding agent access: 83% of orgs run unapproved agents, 99% of repo tokens never expire
Matéria consolidada. Reúne 2 matérias que o Tifon tinha publicado separadamente sobre a mesma história. A IA reescreveu o relatório com o que a fonte trouxe em 06/11/2026 12:15; os links antigos trazem para cá.
Fontes: helpnetsecurity.com ×2 ↗
BlueFlag Security published vendor research, described as The State of Developer Risk 2026 (helpnetsecurity.com, 2026-11-06), showing widespread excess access in software development environments. The findings cover developers, non-human identities and AI coding agents. An earlier article (helpnetsecurity.com, 2026-11-04) covered the company's expanded AI Agent Governance platform and cited the same 83% figure for unapproved AI coding agents. This is vendor research on pooled data, not a disclosed incident or vulnerability. No threat actor or CVE is named. Key findings: 99% of measured repository-access tokens never expire, and agents often run under administrator-level identities with no behavioral monitoring.
Detalhes técnicos
- Unapproved AI coding agents were present in 83% of organizations, and most had no list of approved tools (helpnetsecurity.com). A typical organization used about seven AI coding tools, and some used more than 30.
- 56% of organizations have autonomous agents operating under their own identity. In more than 1 in 5 organizations an agent has merged code with zero human review (helpnetsecurity.com, 2026-11-04).
- Administrator permissions were held by 68% of the identities that agents operated under, and 81% of those identities lacked behavioral monitoring (helpnetsecurity.com, 2026-11-06). In six in ten organizations with AI activity, agents used tokens capable of cloning every repository.
- The average developer had not used 61% of their granted permissions. Among the developers with the most unused permissions, 88% could reach repositories classified as business-critical.
- Four in five organizations had developer identities inactive for more than 90 days that still held administrator permissions.
- Non-human identities with repository access outnumbered human developers eight to one, counting access tokens as identities.
- Of the repository-access tokens measured, 99% had no expiration date, 84% allowed code changes and 47% had been unused for more than 90 days without being revoked.
- Half of the organizations had at least one identity with a spike in code cloning. In 28%, an identity cloned at least 100 repositories it had never accessed. The report says this warrants investigation but does not prove credential misuse or code theft.
- Four in ten repositories had unprotected main branches, and administrators bypassed code review in four in five organizations. Code was merged despite failed checks in 56% of organizations and with unresolved reviewer comments in two-thirds.
- Private repositories became public in 28% of organizations, which can expose code, documentation and stored credentials.
- The unmanaged attack surface is described as invisible to code scanning tools and traditional identity management platforms (helpnetsecurity.com, 2026-11-04).
- The AI Agent Governance platform adds six capabilities (helpnetsecurity.com, 2026-11-04):
- AI agent classification and approval tracking.
- Shadow AI detection on developer endpoints through CrowdStrike and SentinelOne EDR integration, with no new endpoint agents.
- Human-in-the-loop checks for agent-approved pull requests, fully agent-written commits and unreviewed pipeline deployments.
- An agentic application inventory assessed against the OWASP Agentic Top 10.
- AI spend-to-output tracking.
- Two remediation tools: AccessGuard for identity hygiene and DriftGuard for pipeline and repository misconfiguration drift, such as missing branch protections.
Remediação
Neither article gives formal remediation steps for a specific flaw. The findings point to these controls:
- Inventory all AI coding agents, approved and shadow, and keep a list of approved tools.
- Review access regularly and remove unused permissions, stale credentials and dormant memberships.
- Expire or revoke stale tokens.
- Monitor agent identities with the same rigor as human and machine identities.
- Protect main branches and enforce human code review before merge or deploy of agent-generated code.
Relatos divergentes
The two articles name the research differently: 2026 State of Developer Identity Risk (helpnetsecurity.com, 2026-11-04) and The State of Developer Risk 2026 (helpnetsecurity.com, 2026-11-06). The sources do not say whether these are one report or two. The shared 83% figure is consistent. The 2026-11-06 article is the more recent and more detailed.
Linha do tempo
-
BlueFlag report: 99% of repository tokens never expire and AI coding agents hold excess privileges
ler a análise desta fonte ocultar
BlueFlag Security's The State of Developer Risk 2026 report describes widespread excess access in software development environments. It covers developers, non-human identities and AI coding agents. It is vendor research on pooled data, not a specific incident or vulnerability. It names no threat actor and no CVE.
Technical Details
- The average developer had not used 61% of their granted permissions. Among the developers with the most unused permissions, 88% could reach repositories classified as business-critical.
- Four in five organizations had developer identities inactive for more than 90 days that still held administrator permissions.
- Non-human identities with repository access outnumbered human developers eight to one, counting access tokens as identities.
- Of the repository-access tokens measured, 99% had no expiration date, 84% allowed code changes and 47% had been unused for more than 90 days without being revoked.
- A typical organization used about seven AI coding tools, and some used more than 30. Unapproved agents were present in 83% of organizations, and most had no list of approved tools.
- Administrator permissions were held by 68% of the identities that agents operated under, and 81% of those identities lacked behavioral monitoring. In six in ten organizations with AI activity, agents used tokens capable of cloning every repository.
- Half of the organizations had at least one identity with a spike in code cloning, and in 28% an identity cloned at least 100 repositories it had never accessed. The report says this warrants investigation but does not prove credential misuse or code theft.
- Four in ten repositories had unprotected main branches, and administrators bypassed code review in four in five organizations. Code was merged despite failed checks in 56% of organizations and with unresolved reviewer comments in two-thirds.
- Private repositories became public in 28% of organizations, which can expose code, documentation and stored credentials.
Remediation
The article gives no explicit remediation steps. The findings point to these controls:
- Review access regularly and remove unused permissions.
- Expire or revoke stale tokens.
- Keep an inventory of approved AI tools.
- Monitor the behavior of agent identities.
- Protect main branches and enforce code review.
-
BlueFlag Security Launches AI Agent Governance Platform, Reports 83% of Orgs Run Unapproved AI Coding Agents
ler a análise desta fonte ocultar
BlueFlag Security has expanded its AI Agent Governance platform to map access, behavior and oversight of approved and shadow AI agents across the software development lifecycle (SDLC), including automated risk remediation. The announcement follows BlueFlag's 2026 State of Developer Identity Risk report, which found that 83% of organizations run AI coding agents that were never formally approved, 56% have autonomous agents operating under their own identity, and in more than 1 in 5 organizations an agent has merged code with zero human review. Many agents operate with administrator-level permissions under developer identities without any monitoring, creating significant unmanaged attack surface in SDLC environments that is invisible to code scanning tools and traditional identity management platforms. This is a vendor product announcement rather than a disclosed incident, but the underlying survey data highlights a systemic governance gap in enterprise AI adoption.
Technical Details
The platform adds six capabilities: AI agent classification and approval tracking (mapping commits, pull requests, and repositories tied to each agent); shadow AI detection on developer endpoints via integration with CrowdStrike and SentinelOne EDR platforms without requiring new endpoint agents; human-in-the-loop checks flagging agent-approved pull requests, fully agent-written commits, and unreviewed pipeline deployments; an agentic application inventory assessing model providers and frameworks against the OWASP Agentic Top 10; AI spend-to-output tracking; and two remediation tools — AccessGuard for identity hygiene (over-permissioned identities, stale credentials, dormant memberships) and DriftGuard for pipeline/repository misconfiguration drift (e.g., missing branch protections).
Remediation
Organizations are encouraged to inventory all AI coding agents (approved and shadow), enforce human review gates before merge/deploy for agent-generated code, monitor agent identities with the same rigor as human and machine identities, and remediate over-permissioned or stale credentials tied to AI agents.