AI 架构治理文章分析工具调用中途撤销权限的生效机制

On November 4, 2026, the AI section of DEV Community published an article titled "AI architecture governance: what happens when permissions are revoked mid-tool-call," laying out a clear claim about when revocation takes effect. According to the article, once a tool call has passed through the gateway and is still running, revoking the role's permissions does not interrupt the call in flight — it keeps running and finishes. Revocation governs the next call, not the current one. The author is careful to separate this from the design intent: this behavior is not what the design intended, it is simply what the current implementation does. The article then lays out three follow-up threads of analysis. First, what each of the three undo acts actually stops. Second, at what point the gateway takes effect. Third, what would have to be added to make a running call stoppable. At the top of the piece the author adds a note that the first sentence may be all some readers need, and that the remainder can be skipped, which points to the article's main finding being front-loaded before the detailed walkthrough of the three undo actions, the gateway timing, and the missing mechanisms.