Is MONARK MCP server safe?

Findings (1)

• lowUnusually long tool description (over 2,000 characters)poison.long-description

  tool gate: …Coverage-gated decision from the real HIKAE L3 policy (commit/defer/abstain) over a caller-carried authorization budget B_t. Dispatches on task_class. For 'btc-dir-15m' it conformalizes against a committed synthetic calibration derived from the HIKAE S2a instrument (seed 101, n=300 draw), declared synthetic — a plumbing fixture, not a measured predictor. For 'cascade-liquidable-24h' no cascade calibration is committed; the gate abstains (under_calib) on this class. For 'stable-run-velocity-24h' (Narabi: a redemption-flow velocity forecast) the gate holds a committed stable-run velocity calibration for the USDe synthetic-dollar-whitelisted-redeem population (key narabi:persistence-v2@eip155:1/erc20:0x4c9edd5852cd905f086c759e8383e09bff1e68b3) over calm-window redemption flow; coverage is stated under the split-conformal bound of Barber, Candes, Ramdas and Tibshirani 2023 (Thm 2, unit weights): at least 1 − α minus the average total-variation gap between calibration windows and the next one; that gap is not estimated here and the calibration is measured non-stationary across half-years, so 1 − α is the coverage only if that gap is zero (exchangeability), which is not assumed here; no coverage is measured; every other (task_class, predictor_id) abstains (under_calib); for any other population, no stable-run velocity calibration is committed for this population; the gate abstains (under_calib). When the caller instead supplies a `calibration` (its own nonconformity scores plus a `mode`: `interval` ⇒ region [yhat - q̂, yhat + q̂], or `set` ⇒ a conformal set over caller `candidates`), the gate conformalizes against THOSE caller-supplied scores (BYO): split-conformal quantile at miscoverage α over caller-supplied nonconformity scores. MONARK does not see, store, or verify the caller's data or model, and does not validate that the supplied numbers are nonconformity scores of any model. Marginal 1−α coverage holds ONLY for future points exchangeable with the supplied scores; non-exchangeable data (e.g. distribution-shifted or time-ordered) voids it. Never a probability of being right. The gate only emits a decision; it never calls the named tool.…