CVE-2026-115477: obsidian-tc: HITL approval bypass runs destructive tools

obsidian-tc is a governed, agent-ready Model Context Protocol server for Obsidian vaults. From 1.13.0 until 1.31.4, on a Streamable HTTP deployment with `auth.jwtSecret` configured and a 2026-07-28-era client declaring `elicitation.form` support, a dispatch-gated tool call offered an `inputRequired` confirmation could still complete after the human declined or cancelled, or the elicitation response was omitted or malformed, provided the client echoed the validly bound `requestState` on resubmission. The server trusted an authentic, call-bound state as proof of approval even though that state was minted before the `elicitation/create` leg was asked and proved only that the resubmission belonged to the same call. A client behaving incorrectly or under an attacker's control could therefore execute dispatch-gated destructive vault operations without the human's approval, while still requiring the caller's ordinary tool permissions. Only tools marked `destructive: true` or requiring a scope with a human-in-the-loop floor were reachable through this bypass. The sixteen handler-side conditionally gated tools, including `write_note` overwrite and `move_note`, did not read elicitation state on the modern wire and let neither approval nor decline through that path. Affected deployments using stdio or leaving `auth.jwtSecret` unset instead used the CLI `elicit_token` fallback and did not offer this `requestState` round trip. The fix does not prevent a malicious client from reporting a fabricated approval, because the server cannot verify that the client displayed the form or that a human saw it. This issue is fixed in version 1.31.4.

An attacker controls, or has tampered with, an MCP client or agent harness that talks to a shared obsidian-tc server over Streamable HTTP. The client calls a destructive tool, receives the inputRequired response with a bound requestState, and then resubmits with the state, either without showing the confirmation form or after the user declines. The server treats the authentic state as proof of approval and runs the destructive operation. The attacker can also combine this with a prompt injection that steers an agent into a destructive tool call, since the human decline no longer blocks execution. The result is deletion or corruption of notes in the vault without the required human consent.