Engineering should treat the next deploy as a trust-path decision, not just a patching queue. I’d freeze only the build paths that can execute untrusted code: npm installs, CI jobs pulling public actions/plugins, Gitea-hosted repos, Jenkins/GitHub Actions workflows, container builds, and any AI framework package updates. Do not freeze all production releases indefinitely; freeze dependency upgrades and rebuilds until you can prove the exact package versions, build runners, and registry tokens were not exposed.
For the npm worm, the visible evidence says this is publisher-credential compromise plus automated republishing: reports cite 444 packages, 2,000+ versions, stolen npm/GitHub/cloud/Kubernetes/Vault credentials, and preinstall execution. That means engineering should immediately block known poisoned versions, disable lifecycle scripts during install where possible, rebuild CI runners from clean images, rotate npm publish tokens, GitHub Actions secrets, cloud keys, Kubernetes configs, Vault tokens, SSH deploy keys, and any tokens reachable from developer workstations that installed affected versions. Rotation without rebuild is weak; if the worm ran in the environment, assume it read the old secrets.
For Gitea, the source pack has CVE-2026-59774 as actively exploited, but the visible detail is limited. The supply-chain risk is repository and CI trust: private code, deploy keys, workflow files, and release credentials may be exposed if the instance was reachable. Before deploy: patch or isolate Gitea, review new/modified hooks, release tags, CI configs, deploy keys, OAuth apps, and commits pushed during the exposure window. For Paperclip CVE-2026-41679 and the CrowdStrike DPRK AI-package claim, I don’t have enough visible source detail here to scope exact package names or affected versions, so I would not order a blanket AI stack purge. I would instead freeze AI framework package updates, require hash-pinned installs, verify package provenance, and compare SBOMs against the vendor’s affected package list once available.
Provenance checks before the next deploy should be mechanical: generate an SBOM for the release candidate, diff it against the last known-good build, fail the build on unexpected new maintainers/packages/version jumps, require lockfiles with integrity hashes, prefer signed artifacts/Sigstore where available, and verify GitHub Actions/Jenkins plugins are pinned by commit or immutable version rather than floating tags. Where not to overreact: don’t rotate every enterprise credential if it was never present in developer/CI environments; don’t rebuild production hosts that only consumed already-built internal artifacts; and don’t block emergency business fixes if their dependency graph is unchanged and built in a freshly provisioned, token-rotated runner.