Case · DiggingBeagle record

wfx-loot used compromised GitHub identities to stage CI credential harvesting, including AI provider keys

The wfx-loot campaign abused already-compromised GitHub accounts to push apparently benign dependency-audit changes and covertly place malicious workflow modifications or package lifecycle hooks. The useful AI-security boundary is the privileged automated build or coding-agent process: ordinary repository material or an npm lifecycle script could run under CI credentials and access AI provider tokens in that environment.

Public disclosure: Nov 12, 2026

Scope

A malicious actor already holds GitHub write authority for a repository or can modify a dependency executed by a trusted workflow.

UnratedAI role: AGAINST AIAssessment method

At a glance

Mechanism and trust boundary

  1. 01

    Starting condition

    A malicious actor already holds GitHub write authority for a repository or can modify a dependency executed by a trusted workflow.

    Boundary: precondition and attacker capability

  2. 02

    Trust boundary crossed or evaluated

    Some workflow-file changes were held by GitHub approval; a different package.json postinstall route could execute in preexisting trusted workflows. Attackers sought OpenAI, Anthropic, Gemini, Cursor, Copilot, Codex and other model tokens alongside cloud, registry and identity credentials. This is a source-controlled developer automation compromise rather than confirmed autonomous AI-led exploitation.

    Boundary: mechanism

  3. 03

    Observed consequence

    A seemingly documentation-related change mutates CI workflow/installation behavior and transfers environment-level provider credentials to external collection services.

    Boundary: observed effect and limitations

Claims & evidence

CLM-NOV15-WFXLOOT-MECHThe wfx-loot campaign abused already-compromised GitHub accounts to push apparently benign dependency-audit changes and covertly place malicious workflow modifications or package lifecycle hooks. The useful AI-security boundary is the privileged automated build or coding-agent process: ordinary repository material or an npm lifecycle script could run under CI credentials and access AI provider tokens in that environment. Some workflow-file changes were held by GitHub approval; a different package.json postinstall route could execute in preexisting trusted workflows. Attackers sought OpenAI, Anthropic, Gemini, Cursor, Copilot, Codex and other model tokens alongside cloud, registry and identity credentials. This is a source-controlled developer automation compromise rather than confirmed autonomous AI-led exploitation.supported

Basis: reported finding

Link to claim
CLM-NOV15-WFXLOOT-SCOPEThe Aikido public report counted 4,837 compromised accounts and 61,764 targeted repositories, while other researchers inspected overlapping narrower samples. The repo and account census must not be equated with executed hooks, delivered secrets or individual victims. Controlled postinstall evidence should be separated from general probe counts. No AI model is shown choosing the attack. The AI-specific finding concerns credential reuse and automation permissions available when a coding-agent or CI tool consumes attacker-modified files.supported

Basis: reported finding

Link to claim

Implications

A seemingly documentation-related change mutates CI workflow/installation behavior and transfers environment-level provider credentials to external collection services.

Controls and mitigations

  • Enforce independent approval for changed workflows and dependency installation scripts; prevent untrusted PR code from accessing provider tokens.
  • Rotate model/cloud keys exposed to compromised CI jobs, inspect OIDC and federated-identity persistence, and distinguish held workflows from postinstall execution.

Unknowns and contradictions

  • No AI model is shown choosing the attack. The AI-specific finding concerns credential reuse and automation permissions available when a coding-agent or CI tool consumes attacker-modified files.

Sources and citation

Material revision history

  1. Nov 15, 2026 · Canonical change recorded · new in release · revision 250