feat: flag glyph covert-channel risks

src/core/unicode-safety.test.ts

@@ -1,5 +1,5 @@
 import { describe, expect, it } from "vitest";
-import { findHiddenUnicode, hasHiddenUnicode, stripHiddenUnicode } from "./unicode-safety.js";
+import { assessLegibilityRisk, findHiddenUnicode, hasHiddenUnicode, stripHiddenUnicode } from "./unicode-safety.js";
 
 describe("unicode safety", () => {
   it("strips hidden instruction carriers", () => {
@@ -10,4 +10,17 @@ describe("unicode safety", () => {
     expect(stripHiddenUnicode(text)).toBe("safetext");
     expect(findHiddenUnicode(text).map((hit) => hit.codePoint)).toEqual(["U+E0069", "U+200B", "U+202E"]);
   });
+
+  it("flags private-use glyph covert channels", () => {
+    const risk = assessLegibilityRisk(`function ${String.fromCharCode(0xe000)}() { return ${String.fromCharCode(0xe001)}; }`);
+
+    expect(risk.risky).toBe(true);
+    expect(risk.reason).toContain("private-use");
+  });
+
+  it("allows ordinary non-English text", () => {
+    const risk = assessLegibilityRisk("Привет мир, обычный текст без кода");
+
+    expect(risk.risky).toBe(false);
+  });
 });

src/core/unicode-safety.ts

@@ -1,4 +1,13 @@
 const HIDDEN_UNICODE = /[\u200B-\u200F\u202A-\u202E\u2060-\u206F\uFE00-\uFE0F\u{E0000}-\u{E007F}]/gu;
+const CODE_LIKE = /\b(function|const|let|var|class|import|export|return|if|for|while|=>|docker|curl|bash|npm|node|python)\b|[{}();=<>]/u;
+
+export interface LegibilityRisk {
+  risky: boolean;
+  reason: string;
+  nonAsciiRatio: number;
+  symbolRatio: number;
+  privateUseCount: number;
+}
 
 export function stripHiddenUnicode(text: string): string {
   // ponytail: strips known invisible instruction carriers; add reporting metadata if UI needs forensics.
@@ -17,3 +26,22 @@ export function findHiddenUnicode(text: string): { index: number; codePoint: str
     codePoint: `U+${match[0].codePointAt(0)!.toString(16).toUpperCase().padStart(4, "0")}`,
   }));
 }
+
+export function assessLegibilityRisk(text: string): LegibilityRisk {
+  const chars = [...stripHiddenUnicode(text)].filter((c) => !/\s/u.test(c));
+  if (chars.length === 0) return { risky: false, reason: "empty", nonAsciiRatio: 0, symbolRatio: 0, privateUseCount: 0 };
+  const nonAscii = chars.filter((c) => c.codePointAt(0)! > 0x7f).length;
+  const symbols = chars.filter((c) => /[\p{S}\p{Co}]/u.test(c)).length;
+  const privateUseCount = chars.filter((c) => /\p{Co}/u.test(c)).length;
+  const nonAsciiRatio = nonAscii / chars.length;
+  const symbolRatio = symbols / chars.length;
+  const codeLike = CODE_LIKE.test(text);
+  // ponytail: conservative covert-channel tripwire; tune only with real false positives.
+  const risky = privateUseCount > 0 || (codeLike && chars.length >= 20 && nonAsciiRatio > 0.35 && symbolRatio > 0.2);
+  const reason = privateUseCount > 0
+    ? "private-use glyphs present"
+    : risky
+      ? "code-like text has low human legibility"
+      : "legible enough";
+  return { risky, reason, nonAsciiRatio, symbolRatio, privateUseCount };
+}

src/index.ts

@@ -3072,6 +3072,7 @@ security
   .option("--include-fixtures", "Include intentional red-team fixtures/generators")
   .action(async (target: string, opts: { includeFixtures?: boolean }) => {
     const { findPromptInjection } = await import("./core/prompt-injection-safety.js");
+    const { assessLegibilityRisk } = await import("./core/unicode-safety.js");
     const root = path.resolve(target);
     const files: string[] = [];
 
@@ -3094,10 +3095,13 @@ security
     walk(root);
     let hits = 0;
     for (const file of files) {
-      const found = findPromptInjection(fs.readFileSync(file, "utf-8")).filter((h) => h.kind !== "instruction-smuggling");
+      const content = fs.readFileSync(file, "utf-8");
-      if (found.length === 0) continue;
+      const found = findPromptInjection(content).filter((h) => h.kind !== "instruction-smuggling");
-      hits += found.length;
+      const legibility = assessLegibilityRisk(content);
-      console.log(`${file}: ${found.map((h) => `${h.kind}:${h.match}@${h.index}`).join(", ")}`);
+      if (found.length === 0 && !legibility.risky) continue;
+      hits += found.length + (legibility.risky ? 1 : 0);
+      if (found.length > 0) console.log(`${file}: ${found.map((h) => `${h.kind}:${h.match}@${h.index}`).join(", ")}`);
+      if (legibility.risky) console.log(`${file}: human-legibility-risk:${legibility.reason}`);
     }
 
     if (hits > 0) {

src/upgrades/content-safety-gate.test.ts

@@ -55,6 +55,13 @@ describe("ContentSafetyGate", () => {
     expect(v.action).toBe("allow");
   });
 
+  it("blocks human-unreadable glyph code", () => {
+    const glyph = String.fromCharCode(0xe000);
+    const v = gate.evaluate(`function ${glyph}() { return ${glyph}; }`);
+    expect(v.action).toBe("block");
+    expect(v.classification.reason).toContain("human-legibility");
+  });
+
   it("reformulates medium-risk financial advice", () => {
     const v = gate.evaluate("Give me stock trading advice");
     expect(["reformulate", "block", "allow"]).toContain(v.action);

src/upgrades/content-safety-gate.ts

@@ -1,3 +1,5 @@
+import { assessLegibilityRisk } from "../core/unicode-safety.js";
+
 /**
  * Content Safety Gate — classifies tasks by risk domain before they reach the loop.
  *
@@ -139,6 +141,17 @@ export class ContentSafetyGate {
       };
     }
 
+    const legibility = assessLegibilityRisk(task);
+    if (legibility.risky) {
+      return {
+        domain: "cybersecurity_exploit",
+        level: "high",
+        confidence: 0.9,
+        reason: `Task has human-legibility/covert-channel risk: ${legibility.reason}`,
+        reformulation: "Provide reviewable ASCII/source text before requesting execution, deployment, or code review.",
+      };
+    }
+
     // Check each domain for matches
     for (const [domain, patterns] of this.domainPatterns) {
       for (const pattern of patterns) {