GHAgentCI mined historical GitHub workflow revisions, then classified agent invocations and visible guard settings at the file level. It distinguishes a guard absent from a workflow file from a guard actually absent at runtime, since platform and organization defaults may enforce additional restrictions. The dataset reports 159,878 historical revisions and…
Inspect the ClaimsCase · DiggingBeagle record
Most sampled coding-agent CI workflows lacked at least one explicitly declared permission or artifact-pinning control
GHAgentCI mined historical GitHub workflow revisions, then classified agent invocations and visible guard settings at the file level. It distinguishes a guard absent from a workflow file from a guard actually absent at runtime, since platform and organization defaults may enforce additional restrictions.
Public disclosure: Nov 11, 2026
Coding agents are triggered in CI by issues, pull requests and other events across repositories of varying trust.
At a glance
An agent action runner inherits whatever effective token, workflow trigger, installation and network capabilities remain at execution, not merely what the YAML appears to request.
Read the implicationsThe evidence is bounded by the cited source and test setting. Broader incidence, other environments, and actual production consequences are not established merely by these observations.
Limits and uncertaintyFull account
GHAgentCI mined historical GitHub workflow revisions, then classified agent invocations and visible guard settings at the file level. It distinguishes a guard absent from a workflow file from a guard actually absent at runtime, since platform and organization defaults may enforce additional restrictions.
The dataset reports 159,878 historical revisions and 9,265 agentic workflows in the latest snapshot. Explicit token permissions appeared in roughly 19 percent and dedicated action SHA pinning in 24 percent. This is a reproducible configuration census, not an observed exploitation rate.
The sample cannot establish effective permissions for every repository or the outcome of tool calls; deriving compromise incidence from missing YAML declarations would exceed the evidence.
Reported figures and denominator. Latest 9,265 workflows: explicit token permissions around 19%; dedicated-action SHA pins around 24%; neither number is attack incidence.
The evidence is bounded by the cited source and test setting. Broader incidence, other environments, and actual production consequences are not established merely by these observations.
Mechanism and trust boundary
- 01
Starting condition
Coding agents are triggered in CI by issues, pull requests and other events across repositories of varying trust.
Boundary: precondition and attacker capability
- 02
Trust boundary crossed or evaluated
The dataset reports 159,878 historical revisions and 9,265 agentic workflows in the latest snapshot. Explicit token permissions appeared in roughly 19 percent and dedicated action SHA pinning in 24 percent. This is a reproducible configuration census, not an observed exploitation rate.
Boundary: mechanism
- 03
Observed consequence
An agent action runner inherits whatever effective token, workflow trigger, installation and network capabilities remain at execution, not merely what the YAML appears to request.
Boundary: observed effect and limitations
Claims & evidence
CLM-NOV15-GHAGENTCI-MECHGHAgentCI mined historical GitHub workflow revisions, then classified agent invocations and visible guard settings at the file level. It distinguishes a guard absent from a workflow file from a guard actually absent at runtime, since platform and organization defaults may enforce additional restrictions. The dataset reports 159,878 historical revisions and 9,265 agentic workflows in the latest snapshot. Explicit token permissions appeared in roughly 19 percent and dedicated action SHA pinning in 24 percent. This is a reproducible configuration census, not an observed exploitation rate.supported
Basis: reported finding
- supportsGHAgentCI: Coding-Agent Invocations in CI Workflow Historiesprimary disclosure
GHAgentCI dataset, reported 159878 revisions, latest workflow guard rates and manual validation
CLM-NOV15-GHAGENTCI-SCOPEThe sample cannot establish effective permissions for every repository or the outcome of tool calls; deriving compromise incidence from missing YAML declarations would exceed the evidence. The evidence is bounded by the cited source and test setting. Broader incidence, other environments, and actual production consequences are not established merely by these observations.supported
Basis: reported finding
- supportsGHAgentCI: Coding-Agent Invocations in CI Workflow Historiesprimary disclosure
Primary disclosure: results, scope and limitations
Implications
An agent action runner inherits whatever effective token, workflow trigger, installation and network capabilities remain at execution, not merely what the YAML appears to request.
Controls and mitigations
- Inventory actual effective GITHUB_TOKEN and installation scopes at runtime, including pull_request_target and reusable-workflow inheritance.
- Pin action code and restrict comment/request triggers to authorized actors, then verify outcomes with run logs and effect receipts.
Unknowns and contradictions
- The evidence is bounded by the cited source and test setting. Broader incidence, other environments, and actual production consequences are not established merely by these observations.
Sources and citation
Material revision history
- Nov 15, 2026 · Canonical change recorded · new in release · revision 250