Score your CI/CD pipeline against the four preconditions of agentjacking. Get a defensive re-scoping checklist for your next audit.
The June disclosures named a failure class: AI coding agents acting on attacker-controlled text. The governance question is whether an agent with repo-write and secrets-read access is a productivity tool or a privileged identity. Four preconditions have to co-occur before agentjacking becomes a live exposure. Check the ones that describe your pipeline. The full case lives in Agentjacking Goes Public.
Interactive Widget · Agent Privilege Scoreboard
Check each precondition that applies to your AI coding agent.
The four boxes below are the failure class. One alone is manageable. Four together is agentjacking-ready — the state Anna's June briefing warned about.
Preconditions · Check All That Apply
Precondition 1
Untrusted metadata ingestion
Does your agent ingest text from GitHub issues, PR descriptions, calendar invites, contact fields, or any user-controllable metadata?
Precondition 2
Repo write scope
Does your agent have write access to any repository, branch, or artifact registry?
Precondition 3
Secrets read scope
Does your agent have access to OIDC tokens, CI environment variables, service-account credentials, or vault secrets during execution?
Precondition 4
No per-action approval
Does your agent execute privileged actions (writes, secret reads, deploys) without explicit per-action human approval?
0/4
Risk Tier
Green
Standard access controls apply. No urgent action needed on the agentjacking failure class.
Mitigation Checklist
Standard cadence
The precondition framing comes from the June 2026 briefing on the OpenClaw, Claude Code, and npm-supply-chain disclosures. This scoreboard is a starting checklist, not an audit substitute. Treat any Amber or Red result as a trigger for a re-scoping conversation with your security and procurement leads.