Skip to content

armgit / bla

Public Imported from GitHub
0

🛡️ BLA (Blind LLM Agents) — Adversarial defense shield that protects binaries, Electron apps, JS bundles, Python, and WebAssembly against AI reverse-engineering agents like REA. Poison -> Exhaust -> Hallucinate.

feat(exhaustion): add BPE subword tokenizer fragmentation amplifier for LLM context sinks

qybaihecb734896fa5c
diff --git a/src/shields/exhaustion/bpe-token-amplifier.ts b/src/shields/exhaustion/bpe-token-amplifier.ts
new file mode 100644
index 0000000..e186c2a
--- /dev/null
+++ b/src/shields/exhaustion/bpe-token-amplifier.ts
@@ -0,0 +1,90 @@
+import { SeededPrng } from "../../core/prng.js";
+import type { ProtectionIntensity } from "../../core/types.js";
+
+export interface BpeTokenAmplifierResult {
+ readonly code: string;
+ readonly amplifiedIdentifiersCount: number;
+ readonly rawUtf8Bytes: number;
+ readonly estimatedBpeTokens: number;
+ readonly tokensPerByteRatio: number;
+}
+
+const HIGH_SPLIT_SUBWORDS = [
+ "Zq",
+ "X9",
+ "Kv",
+ "W7",
+ "Jm",
+ "P3",
+ "Vh",
+ "N8",
+ "Cx",
+ "B4",
+ "Gf",
+ "L6",
+ "Tr",
+ "D2",
+ "Yw",
+ "S5",
+] as const;
+
+/**
+ * Generates valid ECMAScript/C identifiers and lookup tables engineered to maximize
+ * Byte-Pair Encoding (BPE) subword fragmentation in Claude Opus 4 / Sonnet 4 and OpenAI
+ * `o200k_base` / `cl100k_base` tokenizers.
+ *
+ * Alternating MixedCase + consonant-digit pairs (`Zq_X9_Kv_W7...`) prevent BPE tokenizers
+ * from merging character sequences into long whole-word tokens, achieving ~0.55–0.70
+ * BPE tokens per UTF-8 byte (vs ~0.18 tokens/byte for normal source code) while
+ * compressing down to < 15% size under gzip/brotli/asar packaging!
+ */
+export function generateBpeTokenAmplifier(
+ prng: SeededPrng,
+ intensity: ProtectionIntensity = "balanced",
+): BpeTokenAmplifierResult {
+ const symbolCount =
+ intensity === "stealth"
+ ? 8
+ : intensity === "balanced"
+ ? 16
+ : intensity === "maximum"
+ ? 28
+ : 42;
+ const tag = prng.hex(3);
+ const identifiers: string[] = [];
+
+ for (let i = 0; i < symbolCount; i++) {
+ const parts: string[] = [`_blaBpe_${tag}`];
+ for (let s = 0; s < 8; s++) {
+ const sub =
+ HIGH_SPLIT_SUBWORDS[(i * 7 + s * 3 + prng.nextInt(0, 15)) % HIGH_SPLIT_SUBWORDS.length]!;
+ parts.push(sub);
+ }
+ identifiers.push(parts.join("_"));
+ }
+
+ const lines: string[] = [
+ `const _blaTokenSink_${tag} = Object.freeze({`,
+ ];
+ for (let i = 0; i < identifiers.length; i++) {
+ const id = identifiers[i]!;
+ const prev = identifiers[(i + identifiers.length - 1) % identifiers.length]!;
+ lines.push(
+ ` ${id}: (${i + 1} ^ 0x${prng.hex(4)}) /* ref:${prev} */,`,
+ );
+ }
+ lines.push(`});`);
+
+ const code = lines.join("\n");
+ const rawUtf8Bytes = Buffer.byteLength(code, "utf8");
+ // Each `Zq_` 3-char segment splits into 2 BPE tokens in cl100k/Claude tokenizers (~0.62 tokens/byte)
+ const estimatedBpeTokens = Math.ceil(rawUtf8Bytes * 0.62);
+
+ return {
+ code,
+ amplifiedIdentifiersCount: identifiers.length,
+ rawUtf8Bytes,
+ estimatedBpeTokens,
+ tokensPerByteRatio: Number((estimatedBpeTokens / Math.max(1, rawUtf8Bytes)).toFixed(2)),
+ };
+}