CVE-2026-58439

CVE-2026-58439 is a state invalidation failure in pull request approval workflows. When a PR's target branch changes, existing approval flags are not recomputed against the new branch's protection rules. An approval granted for `feature-branch` persists unchanged when the PR is retargeted to `main`, and branch protection evaluates that persisted approval against its own criteria without detecting the context shift. The approval survives the retarget operation as a durable artifact while the protection layer treats it as contextually valid — these are incompatible assumptions about what an approval means, and neither system validates the mismatch. This is not a simple injection or privilege escalation bug. It's a boundary contract failure between two systems with different mental models: the approval workflow treats approvals as durable objects (reasonable UX — you don't want approvals evaporating on rebase), while branch protection treats them as context-sensitive signals (correct security posture — merging to `main` should require fresh approval). No interface exists between these systems to negotiate semantic validity. The protection layer accepts approval objects without independently validating their target context, trusting the approval layer to have invalidated stale decisions. Neither asserts its own invariants. The exploit requires only contributor-level access — any developer with write access can retarget a PR and merge without the approvals the target branch requires. Automated tooling amplifies this: GitHub Actions that push branches can trigger retargeting, meaning supply chain compromise of any third-party action grants the equivalent of contributor access. Organizations using untrusted Actions may already be vulnerable without active exploitation. Defenders should prioritize three actions. First, audit CI/CD pipelines for any workflows that modify PR target branches and verify whether downstream merge operations re-validate approval against the final target. Second, implement detection rules that flag retarget-then-merge sequences — the window between retargeting and merge is the exploit vector. Third, treat branch protection as a defense-in-depth layer rather than a complete solution; audit what downstream systems (merge queues, deployment triggers, policy engines) trust the approval signal without independent validation. The vulnerability belongs to a recurring authorization pattern where each component assumes security is enforced elsewhere — OAuth token reuse bugs, session invalidation failures, and now approval persistence all follow the same genetic sequence. The fix is not adding a negotiation protocol (which still asks one system to trust another) but making branch protection independently validate target context for every approval it accepts.

Reviewed through automated stages and approved by a human before publication.