agent-sanitizer 2.23.0 → 2.23.2

package/SECURITY.md

@@ -24,6 +24,23 @@ that exposes a secret, or a crash on adversarial input are all in scope. Please
 don't include real credentials in a report—a credential-shaped placeholder is
 enough.
 
+## Dependency supply chain
+
+Dependency versions must age before this repo will resolve them:
+`pnpm-workspace.yaml` sets `minimumReleaseAge` to 4320 minutes (3 days), so a
+package published minutes ago cannot enter the lockfile. The npm compromises
+this defends against are typically detected and unpublished within hours to a
+couple of days, and the window is where that detection happens.
+
+The setting gates resolution only — `pnpm install --frozen-lockfile` installs
+what the lockfile already names — so it constrains `pnpm add` and `pnpm update`,
+which is where a poisoned release would enter.
+
+`minimumReleaseAgeExclude` exempts only `agent-sanitizer` itself, which this
+repository builds and publishes with npm provenance. `tests/test_minimum_release_age.py`
+fails if the window is removed or shortened, if an exemption names a version
+nothing pins, or if a third-party package is exempted.
+
 ## What to expect
 
 The maintainer will acknowledge your report and work with you on a fix and a

package/THREAT-MODEL.md

@@ -35,6 +35,22 @@ table](./README.md#entry-points) maps each to its import.
   Sinhala) or inside an emoji ZWJ sequence. The carve-out fires only when **both**
   neighbors clearly belong to the context, and it is disabled once the total
   invisible count crosses a scatter floor—over-stripping beats under-stripping.
+- **Blank fillers doing real work in their own script**: the Braille blank
+  (U+2800) beside a real cell, a Hangul filler beside a real jamo/syllable. A
+  _run_ of fillers has only fillers for neighbors, so it fails the anchor and is
+  stripped. Because U+2800 _is_ the word space of Unicode Braille and a Hangul
+  filler completes a defective syllable, these are far denser in genuine text
+  than joiners are, so they carry their own document-wide allowance—one
+  preserved blank per two visible anchor-script characters, above a floor—rather
+  than drawing on the joiner/selector preserve budget. The allowance is counted
+  per script (a blank never anchors cross-script, so Korean prose must not fund
+  a Braille channel). Past that ratio no blank of that script is preserved
+  (never half-spaced) and all of them count as payload, which is the density an
+  alternating `syllable filler …` channel needs. Contracted (grade-2) Braille
+  sits closest to the boundary: alphabet wordsigns are single cells, so a
+  passage of mostly one-cell words approaches 1:1 and is stripped like the
+  channel—an accepted residual false positive inherent to a density rule, not a
+  gap, and not worth widening the ratio to reach.
 
 **Reassembly hardening.** The two passes feed each other in _both_ directions:
 stripping an invisible char can reconstitute an ANSI escape its split had

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-sanitizer",
-  "version": "2.23.0",
+  "version": "2.23.2",
   "description": "Defend an agent against hidden-content injection: strip payload-capable invisible Unicode and ANSI, splice out human-invisible HTML, and flag data-exfil URLs in untrusted text before any model sees it.",
   "type": "module",
   "repository": {

package/src/invisible.mjs

@@ -228,9 +228,10 @@ export const CONSECUTIVE_JOINER_CAP = 8;
 // visible characters in a row), exactly like CONSECUTIVE_JOINER_CAP.
 export const CONSECUTIVE_SELECTOR_CAP = 8;
 
-// Floor on the document-wide preserve budget, shared by both preserve kinds
-// (see `kind` in analyzeCarve: joiners AND presentation selectors draw from
-// the same counter). The Joining_Type gate strips joiners that do no
+// Floor on the document-wide preserve budget for joiners, selectors and tag
+// sequences (see `kind` in analyzeCarve — blank fillers are NOT charged here;
+// they have their own allowance, see TOTAL_PRESERVED_BLANK_BUDGET). The
+// Joining_Type gate strips joiners that do no
 // rendering work regardless of count, so the bulk covert channel (ZWNJ
 // scattered through Latin/ASCII/mixed text) is closed by shape, not by
 // counting. What remains is the residual channel of MEANINGFUL joiners/
@@ -260,6 +261,31 @@ export const PRESERVED_JOINER_PER_VISIBLE = 8;
 // text can widen.
 export const PRESERVE_HARD_CAP = 64;
 
+// Floor on the document-wide allowance for PRESERVED blank fillers (the
+// Braille blank and the Hangul fillers — see the blank-filler carve-out). Kept
+// separate from the joiner/selector budget because the two have completely
+// different legitimate densities, and short blank-dense strings (a one-line
+// Braille phrase, a lone archaic syllable) must stay un-clipped.
+export const TOTAL_PRESERVED_BLANK_BUDGET = 16;
+
+// Visible ANCHOR-script code points (a non-blank Braille cell, a non-filler
+// Hangul jamo/syllable) required per preserved blank filler above the floor. A
+// blank is only ever preservable next to one of these, so this ratio is what
+// separates real text from the degenerate channel: U+2800 separates WORDS and a
+// filler completes a syllable, so genuine text spends several anchor characters
+// per blank and stays under one blank per two anchors, while the alternation an
+// attacker needs to stuff a bit per character (`가ᅟ가ᅟ…`, `⠃⠀⠃⠀…`) is exactly
+// 1:1 and fails.
+//
+// Deliberately NOT capped by PRESERVE_HARD_CAP: an absolute ceiling is what
+// truncated long Braille documents, and unlike the joiner channel this one
+// cannot scale on invisible cover text — every additional bit costs the
+// attacker two VISIBLE anchor-script characters, and the blanks themselves
+// render as spacing a reader can see. Over the ratio, NO blank in the document
+// is preserved (all-or-nothing, so a document never comes out half-spaced) and
+// every one of them becomes payload — which then also feeds the scatter floor.
+export const PRESERVED_BLANK_PER_ANCHOR = 2;
+
 // Scripts whose orthography uses ZWNJ/ZWJ between letters as a rendering
 // control. The runtime gate is now script-agnostic (it reads Joining_Type, so it
 // covers every cursive/Brahmic script, not just these), but this list remains
@@ -402,6 +428,17 @@ function isBrahmicConsonantChar(ch) {
 // the anchor and is stripped — the run-length gate falls out of the anchor. The
 // zero-width Mn marks in BLANK_NON_CF (U+034F/17B4/17B5) have no such benign
 // standalone use, so they are never preserved.
+//
+// Blank fillers are NOT charged against the joiner/selector preserve budget:
+// that budget's density model is "~1 preserved invisible per 8 visible chars"
+// (PRESERVED_JOINER_PER_VISIBLE), measured on Persian ZWNJ prose, with a fixed
+// PRESERVE_HARD_CAP ceiling. Blanks are an order of magnitude denser in genuine
+// text — U+2800 IS the word space of Unicode Braille, and a Hangul filler
+// completes a defective syllable — so charging them there mangled real content:
+// a 40-word Braille passage lost 14 of its 39 word spaces (words run together)
+// and a 200-word one kept 64 of 199, with `found` reporting a strip on a
+// perfectly legitimate document. They draw on the anchor-proportional allowance
+// below instead (see TOTAL_PRESERVED_BLANK_BUDGET).
 const BRAILLE_BLANK = 0x2800;
 const HANGUL_FILLERS = new Set([0x115f, 0x1160, 0x3164, 0xffa0]);
 // Code points that trigger the carve-out path for blank fillers (see
@@ -600,6 +637,44 @@ function analyzeCarve(cps) {
     if (isPreservedBlankFiller(cps, i)) return "blank";
     return null;
   });
+  // Blank fillers are budgeted here, document-wide and all-or-nothing, against
+  // the visible anchor-script text rather than against the joiner/selector
+  // counter in carveStrip (see PRESERVED_BLANK_PER_ANCHOR for why the two
+  // cannot share a density model). Deciding it in analyzeCarve rather than in
+  // the emit loop keeps countPayloadInvisible and payloadInvisibleView honest:
+  // a blank the stripper will remove is payload to every consumer, so the
+  // prompt layer's scatter gate sees it without re-deriving the budget.
+  //
+  // Budgeted PER SCRIPT, because a blank never anchors cross-script: pooling the
+  // two anchor counts would let one script's cover text fund the other's
+  // channel, so 400 chars of ordinary Korean prose would buy an unreported
+  // `⠃⠀⠃⠀…` alternation of 200 Braille blanks. The anchor scan is skipped below
+  // the floor: it costs a script regex per visible character, and analyzeCarve
+  // runs on every prompt and tool output.
+  const blankScript = kind.map((k, i) =>
+    k !== "blank"
+      ? null
+      : cps[i].codePointAt(0) === BRAILLE_BLANK
+        ? "braille"
+        : "hangul",
+  );
+  for (const [
+    script,
+    isAnchor,
+  ] of /** @type {[string, (ch: string) => boolean][]} */ ([
+    ["braille", isBrailleCell],
+    ["hangul", isHangul],
+  ])) {
+    const blanks = blankScript.filter((s) => s === script).length;
+    if (blanks <= TOTAL_PRESERVED_BLANK_BUDGET) continue;
+    const anchors = cps.reduce(
+      (n, ch, i) => n + (codes[i] === null && isAnchor(ch) ? 1 : 0),
+      0,
+    );
+    if (blanks > Math.floor(anchors / PRESERVED_BLANK_PER_ANCHOR))
+      for (let i = 0; i < kind.length; i++)
+        if (blankScript[i] === script) kind[i] = null;
+  }
   let payloadInvis = 0;
   let visibleLen = 0;
   for (let i = 0; i < cps.length; i++) {
@@ -790,12 +865,15 @@ function clusterEnds(body) {
 
 /**
  * Carve-out strip (an invisible the carve-out might preserve is present): walk
- * GRAPHEME CLUSTERS, preserving a cluster's joiners/selectors/tags/blank-fillers
- * only where each has its `kind` set AND the text stays under the scatter floor
- * AND the whole cluster fits inside the remaining per-run
- * (CONSECUTIVE_JOINER_CAP / CONSECUTIVE_SELECTOR_CAP) and document-wide
- * (TOTAL_PRESERVED_JOINER_BUDGET) preserve allowance — otherwise every
- * preservable char in that cluster is stripped like any other payload byte.
+ * GRAPHEME CLUSTERS, preserving a cluster's joiners/selectors/tags only where
+ * each has its `kind` set AND the text stays under the scatter floor AND the
+ * whole cluster fits inside the remaining per-run (CONSECUTIVE_JOINER_CAP /
+ * CONSECUTIVE_SELECTOR_CAP) and document-wide (TOTAL_PRESERVED_JOINER_BUDGET)
+ * preserve allowance — otherwise every preservable char in that cluster is
+ * stripped like any other payload byte. Blank fillers are the exception: their
+ * allowance is anchor-proportional and already spent document-wide in
+ * analyzeCarve (see PRESERVED_BLANK_PER_ANCHOR), so here they answer only to
+ * the scatter floor and do not draw on the joiner/selector budget.
  *
  * The budget is charged against the CLUSTER, not the code point, because the
  * cluster is the indivisible unit: charging per code point let a limit fall due
@@ -862,7 +940,10 @@ function carveStrip(body) {
     let joiners = 0;
     let selectors = 0;
     for (let k = start; k < end; k++) {
-      if (kind[k] === null) continue;
+      // "blank" is exempt: analyzeCarve already decided it against the
+      // anchor-proportional allowance, so it neither draws on this budget nor
+      // is stripped by it (only by the scatter floor, via allowCarveOut).
+      if (kind[k] === null || kind[k] === "blank") continue;
       need++;
       if (kind[k] === "joiner") joiners++;
       if (kind[k] === "ivs" || kind[k] === "stdvs") selectors++;
@@ -904,11 +985,13 @@ function carveStrip(body) {
         out += cps[k]; // ordinary visible character
         continue;
       }
-      if (fits && kind[k] !== null) {
+      // A blank filler rides on allowCarveOut alone (its own allowance is
+      // already spent in analyzeCarve); everything else rides on `fits`.
+      if (kind[k] === "blank" ? allowCarveOut : fits && kind[k] !== null) {
         if (kind[k] === "joiner") joinerRun++;
         if (kind[k] === "ivs" || kind[k] === "stdvs") selectorRun++;
-        preservedTotal++;
-        prevVisible = false; // a joiner/selector/tag keeps the cluster open
+        if (kind[k] !== "blank") preservedTotal++;
+        prevVisible = false; // a joiner/selector/tag/blank keeps the cluster open
         out += cps[k];
         continue;
       }

package/types/invisible.d.mts

@@ -102,5 +102,7 @@ export const CONSECUTIVE_SELECTOR_CAP: 8;
 export const TOTAL_PRESERVED_JOINER_BUDGET: 16;
 export const PRESERVED_JOINER_PER_VISIBLE: 8;
 export const PRESERVE_HARD_CAP: 64;
+export const TOTAL_PRESERVED_BLANK_BUDGET: 16;
+export const PRESERVED_BLANK_PER_ANCHOR: 2;
 export const LINGUISTIC_SCRIPTS: string[];
 export { BRAHMIC_CONSONANT_RANGES } from "./joining-type.mjs";