Answer in brief
CVE-2026-44645 records a Medium severity (CVSS 6.5) vulnerability in LiquidJS has a renderLimit DoS guard bypass via empty `{% for %}` body. The current sources do not mark it as known exploited. The current feed maps liquidjs (npm), liquidjs (npm), liquidjs (npm). Check affected ranges and fixed versions before updating.
Analysis pending evidence review
HOL Guard separates source facts from reviewed analysis. See the methodology.
CVSS is 6.5. The current sources do not mark it as known exploited. Treat this as a source-backed prioritization signal, not a statement about your environment.
Analysis status
Analysis pending evidence review
Factual feed record only; HOL analysis is not approved for indexing. Read the methodology.
The current feed maps liquidjs (npm), liquidjs (npm), liquidjs (npm). Check affected ranges and fixed versions before updating.
| Package | Affected range | Fixed version |
|---|---|---|
| liquidjsnpm | <=10.25.7 | Not reported |
| liquidjsnpm | <10.26.0 | 10.26.0 |
| liquidjsnpm | >=0 <10.26.0 | 10.26.0 |
Published upstream
May 27, 2026
Evidence: source:ghsa:source_dates:source-dates:recordSource modified
Oct 27, 2026
Evidence: source:ghsa:source_dates:source-dates:recordFirst seen by HOL
Jun 22, 2026
## Summary The `renderLimit` option — documented in `docs/source/tutorials/dos.md` as the mechanism that "mitigates this by limiting the time consumed by each render() call" — can be fully bypassed by a `{% for %}` (or `{% tablerow %}`) tag whose body is empty. The per-iteration time check is reached only when the body contains at least one template node, so a template like `{%- for i in (1..N) -%}{%- endfor -%}` iterates the full collection without ever consulting `renderLimit`. With a configured `renderLimit` of 50 ms, a single `parseAndRenderSync` call has been observed to consume **2.26 seconds** (~45× over the limit) and scales linearly with `N` up to `memoryLimit`, allowing a low-privileged template author to wedge an event-loop thread for an attacker-chosen duration. ## Details `Render.renderTemplates` is the single point at which `renderLimit` is consulted: ```ts // src/render/render.ts 14: public * renderTemplates (templates: Template[], ctx: Context, emitter?: Emitter): IterableIterator<any> { 15: if (!emitter) { 16: emitter = ctx.opts.keepOutputType ? new KeepingTypeEmitter() : new SimpleEmitter() 17: } 18: const errors = [] 19: for (const tpl of templates) { 20: ctx.renderLimit.check(getPerformance().now()) 21: try { 22: const html = yield tpl.render(ctx, emitter) ... 32: } ``` The check at line 20 lives **inside** the `for (const tpl of templates)` body. When `templates.length === 0`, the loop body never executes, so the limiter is never consulted on that invocation. The `for` tag re-enters `renderTemplates` once per collection item with no independent time check: ```ts // src/tags/for.ts 70: for (const item of collection) { 71: scope[this.variable] = item 72: ctx.continueCalled = ctx.breakCalled = false 73: yield r.renderTemplates(this.templates, ctx, emitter) 74: if (ctx.breakCalled) break 75: scope.forloop.next() 76: } ``` When `{%- for i in (1..N) -%}{%- endfor -%}` is parsed, `this.templates` is `[]`. Each of the `N` calls to `r.renderTemplates(this.templates, ctx, emitter)` therefore performs zero `renderLimit.check()` calls and zero template work — it just spins the JS-level `for` loop and the generator boilerplate. With `N = 30_000_000` this still costs ~2.26 s of CPU, and `N = 100_000_000` costs ~9.6 s, fully bypassing whatever wall-clock budget the integrator configured. The range expression itself is bounded only by `memoryLimit`: ```ts // src/render/expression.ts:67-72 function * evalRangeToken (token: RangeToken, ctx: Context) { const low: number = yield evalToken(token.lhs, ctx) const high: number = yield evalToken(token.rhs, ctx) ctx.memoryLimit.use(high - low + 1) return range(+low, +high + 1) } ``` So the maximum bypass is governed by the (separate) `memoryLimit`, not by `renderLimit`. Integrators following the `docs/source/tutorials/dos.md` guidance — which positions `renderLimit` as the time-based defense — get no time-based defense at all on this code path. ## PoC Reproduced against `[email protected]` (HEAD `34877950`): ```bash # Empty for-body bypasses renderLimit (50 ms) and runs for ~2.26 s: $ node -e "const { Liquid } = require('liquidjs'); const engine = new Liquid({ memoryLimit: 1e9, renderLimit: 50 }); const t = Date.now(); engine.parseAndRenderSync('{%- for i in (1..30000000) -%}{%- endfor -%}', {}); console.log('Took', Date.now()-t, 'ms');" Took 2255 ms # Same template with a single-character body is correctly bounded: $ node -e "const { Liquid } = require('liquidjs'); const engine = new Liquid({ memoryLimit: 1e9, renderLimit: 50 }); try { engine.parseAndRenderSync('{%- for i in (1..30000000) -%}.{%- endfor -%}', {}); } catch(e) { console.log('correctly threw:', e.message); }" correctly threw: template render limit exceeded, line:1, col:1 ``` Scaling `N`: - `N = 30_000_000` → 2255 ms (≈ 45× over the 50 ms limit) - `N = 100_000_000` → 9581 ms (≈ 191× over the 50 ms limit) Time grows linearly with `N`, capped only by `memoryLimit` (default `Infinity`, so the only cap by default is process memory). ## Impact Any liquidjs integrator who follows the upstream DoS guidance and sets a finite `renderLimit` to bound per-render CPU — typical for SaaS / multi-tenant environments where end users author templates (themes, email templates, snippets) — does not get the bound they configured. A single template submission can keep an event-loop thread busy for seconds, which on a Node.js server is sufficient to stall all in-flight requests on that worker. With a large enough range and a permissive `memoryLimit`, the wedge time is attacker-controlled. No data is exposed and no integrity is harmed; impact is availability only. ## Recommended Fix Move the `renderLimit` check to a location that runs unconditionally per `renderTemplates` invocation, so a zero-template body still triggers it; alternatively (or additionally) have iteration tags that invoke `renderTemplates` per element check the limiter themselves once per iteration. ```ts // src/render/render.ts — check at function entry, before the templates loop public * renderTemplates (templates: Template[], ctx: Context, emitter?: Emitter): IterableIterator<any> { if (!emitter) { emitter = ctx.opts.keepOutputType ? new KeepingTypeEmitter() : new SimpleEmitter() } ctx.renderLimit.check(getPerformance().now()) // <-- runs even when templates is empty const errors = [] for (const tpl of templates) { ctx.renderLimit.check(getPerformance().now()) ... } ... } ``` And/or, defensively, in the iteration tags themselves so the guard cost is paid once per element rather than only at re-entry: ```ts // src/tags/for.ts (around line 70) for (const item of collection) { ctx.renderLimit.check(getPerformance().now()) // <-- per-iteration time check scope[this.variable] = item ctx.continueCalled = ctx.breakCalled = false yield r.renderTemplates(this.templates, ctx, emitter) if (ctx.breakCalled) break scope.forloop.next() } // src/tags/tablerow.ts (around line 54) — analogous addition for (let idx = 0; idx < collection.length; idx++, tablerowloop.next()) { ctx.renderLimit.check(getPerformance().now()) ... } ``` The same hardening should be applied anywhere a tag drives an attacker-influenced loop count over a (potentially empty) `templates` array.
Quoted source text, attributed separately from HOL analysis.