mirror of
https://github.com/openclaw/openclaw.git
synced 2026-05-15 13:19:24 +02:00
* fix(exec-approvals): escape control characters in display sanitizers * docs(changelog): add exec approval control-char display sanitizer entry * fix(exec-approvals): redact before escape, cover U+2028/U+2029 in display sanitizers * fix(exec-approvals): strip invisibles before redaction and align forwarder test * fix(exec-approvals): cover Zs bypass and preserve multi-line context on obfuscated secrets * fix(exec-approvals): compare redaction outputs by content, not length * fix(exec-approvals): suppress raw command on bypass; cover non-ASCII Zs in macOS sanitizer * fix(exec-approvals): use position-bitmap bypass detection and bound input size * style(exec-approvals): satisfy oxlint no-new-array-single-argument and SwiftFormat * fix(exec-approvals): iterate by code point and redact before truncating
181 lines
7.9 KiB
TypeScript
181 lines
7.9 KiB
TypeScript
import { redactSensitiveText, resolveRedactOptions } from "../logging/redact.js";
|
|
import type { ExecApprovalRequestPayload } from "./exec-approvals.js";
|
|
|
|
// Escape control characters, Unicode format/line/paragraph separators, and non-ASCII space
|
|
// separators that can spoof approval prompts in common UIs. Ordinary ASCII space (U+0020) is
|
|
// intentionally excluded so normal command text renders unchanged.
|
|
const EXEC_APPROVAL_INVISIBLE_CHAR_REGEX =
|
|
/[\p{Cc}\p{Cf}\p{Zl}\p{Zp}\u00A0\u1680\u2000-\u200A\u202F\u205F\u3000\u115F\u1160\u3164\uFFA0]/gu;
|
|
const EXEC_APPROVAL_INVISIBLE_CHAR_SINGLE =
|
|
/^[\p{Cc}\p{Cf}\p{Zl}\p{Zp}\u00A0\u1680\u2000-\u200A\u202F\u205F\u3000\u115F\u1160\u3164\uFFA0]$/u;
|
|
|
|
// Hard cap on input the sanitizer will process at all. Above this size we return a constant
|
|
// marker without running any regex work, so an attacker cannot force unbounded CPU/memory.
|
|
const EXEC_APPROVAL_MAX_INPUT = 256 * 1024;
|
|
// Soft cap on displayed output. Truncation happens AFTER redaction so a secret near the
|
|
// cutoff is not partially exposed when the cut lands mid-token below a pattern's minimum
|
|
// length (e.g. `ghp_` needs 20+ trailing chars before the `\b` match).
|
|
const EXEC_APPROVAL_MAX_OUTPUT = 16 * 1024;
|
|
const EXEC_APPROVAL_TRUNCATION_MARKER = "…[truncated]";
|
|
const EXEC_APPROVAL_OVERSIZED_MARKER =
|
|
"[exec approval command exceeds display size limit; full text suppressed]";
|
|
|
|
const BYPASS_MASK = "***";
|
|
|
|
function formatCodePointEscape(char: string): string {
|
|
return `\\u{${char.codePointAt(0)?.toString(16).toUpperCase() ?? "FFFD"}}`;
|
|
}
|
|
|
|
function escapeInvisibles(text: string): string {
|
|
return text.replace(EXEC_APPROVAL_INVISIBLE_CHAR_REGEX, formatCodePointEscape);
|
|
}
|
|
|
|
function truncateForDisplay(text: string): string {
|
|
if (text.length <= EXEC_APPROVAL_MAX_OUTPUT) {
|
|
return text;
|
|
}
|
|
return text.slice(0, EXEC_APPROVAL_MAX_OUTPUT) + EXEC_APPROVAL_TRUNCATION_MARKER;
|
|
}
|
|
|
|
// Build a boolean bitmap of positions in `text` that ANY redaction pattern would match.
|
|
// Patterns are applied independently to the raw text (not sequentially against a
|
|
// progressively-redacted view) so later patterns can still find matches that the in-place
|
|
// redaction would have replaced first. That is conservative — it may over-count overlapping
|
|
// matches — but that is acceptable for a coverage check. Indices are UTF-16 code-unit
|
|
// offsets, matching what `matchAll` returns and aligning with `String#length`.
|
|
function computeRedactionBitmap(text: string, patterns: RegExp[]): boolean[] {
|
|
const bitmap: boolean[] = Array.from({ length: text.length }, () => false);
|
|
for (const pattern of patterns) {
|
|
const iter = pattern.flags.includes("g")
|
|
? new RegExp(pattern.source, pattern.flags)
|
|
: new RegExp(pattern.source, `${pattern.flags}g`);
|
|
for (const match of text.matchAll(iter)) {
|
|
if (match.index === undefined) {
|
|
continue;
|
|
}
|
|
const end = match.index + match[0].length;
|
|
for (let i = match.index; i < end; i++) {
|
|
bitmap[i] = true;
|
|
}
|
|
}
|
|
}
|
|
return bitmap;
|
|
}
|
|
|
|
// Iterate by full Unicode code point so astral-plane invisibles (e.g. U+E0061 TAG LATIN
|
|
// SMALL LETTER A, category Cf) are matched as single characters instead of being seen as a
|
|
// surrogate pair whose halves are category Cs and would escape the invisible-char regex.
|
|
function buildStrippedView(original: string): { stripped: string; strippedToOrig: number[] } {
|
|
const strippedChars: string[] = [];
|
|
const strippedToOrig: number[] = [];
|
|
let offset = 0;
|
|
for (const cp of original) {
|
|
if (!EXEC_APPROVAL_INVISIBLE_CHAR_SINGLE.test(cp)) {
|
|
strippedChars.push(cp);
|
|
for (let k = 0; k < cp.length; k++) {
|
|
strippedToOrig.push(offset + k);
|
|
}
|
|
}
|
|
offset += cp.length;
|
|
}
|
|
return { stripped: strippedChars.join(""), strippedToOrig };
|
|
}
|
|
|
|
export function sanitizeExecApprovalDisplayText(commandText: string): string {
|
|
if (commandText.length > EXEC_APPROVAL_MAX_INPUT) {
|
|
// Refuse to display inputs above the hard cap; anything larger must be approved through
|
|
// another channel. Running redaction on a multi-megabyte payload would be a DoS vector.
|
|
return EXEC_APPROVAL_OVERSIZED_MARKER;
|
|
}
|
|
const rawRedacted = redactSensitiveText(commandText, { mode: "tools" });
|
|
const { stripped, strippedToOrig } = buildStrippedView(commandText);
|
|
const strippedRedacted = redactSensitiveText(stripped, { mode: "tools" });
|
|
// Fast path: stripping invisibles did not expose any additional secret-like content, so the
|
|
// raw-view redaction is sufficient. Preserve structure and show invisible-character spoof
|
|
// attempts as `\u{...}` escapes.
|
|
if (strippedRedacted === stripped) {
|
|
return truncateForDisplay(escapeInvisibles(rawRedacted));
|
|
}
|
|
// Detect bypass by position-bitmap coverage. Run each redaction pattern independently on
|
|
// both views and map stripped-view match positions back to original coordinates. If every
|
|
// position the stripped view would mask is also masked by the raw view, the raw view
|
|
// already covered everything — for example, an ordinary multi-line PEM private key where
|
|
// raw produces `BEGIN/…redacted…/END` while stripped collapses to `***`. A real bypass
|
|
// exists only when the stripped view masks at least one original position raw missed (e.g.
|
|
// the tail of an `sk-` token whose prefix-boundary was broken by a spliced zero-width or
|
|
// NBSP character).
|
|
const { patterns } = resolveRedactOptions({ mode: "tools" });
|
|
const rawMask = computeRedactionBitmap(commandText, patterns);
|
|
const strippedMask = computeRedactionBitmap(stripped, patterns);
|
|
let bypassDetected = false;
|
|
for (let i = 0; i < strippedMask.length; i++) {
|
|
if (strippedMask[i] && !rawMask[strippedToOrig[i]]) {
|
|
bypassDetected = true;
|
|
break;
|
|
}
|
|
}
|
|
if (!bypassDetected) {
|
|
return truncateForDisplay(escapeInvisibles(rawRedacted));
|
|
}
|
|
// Bypass path. Project the stripped-view mask back onto original positions, union with the
|
|
// raw-view mask, and emit a rendering where each contiguous masked run becomes a single
|
|
// `***` marker. Invisible characters that fall outside masked runs still render as visible
|
|
// `\u{...}` escapes so multi-line structure and spliced invisibles stay readable. The
|
|
// render loop advances by full code point so astral-plane invisibles are escaped as one
|
|
// `\u{...}` token rather than two separate surrogate escapes (or, worse, passed through
|
|
// unescaped because neither surrogate half matches the Cf regex).
|
|
const unionMask = rawMask.slice();
|
|
for (let i = 0; i < strippedMask.length; i++) {
|
|
if (strippedMask[i]) {
|
|
unionMask[strippedToOrig[i]] = true;
|
|
}
|
|
}
|
|
let out = "";
|
|
let i = 0;
|
|
while (i < commandText.length) {
|
|
if (unionMask[i]) {
|
|
let j = i;
|
|
while (j < commandText.length && unionMask[j]) {
|
|
j++;
|
|
}
|
|
out += BYPASS_MASK;
|
|
i = j;
|
|
continue;
|
|
}
|
|
const codePoint = commandText.codePointAt(i) ?? 0xfffd;
|
|
const cp = String.fromCodePoint(codePoint);
|
|
out += EXEC_APPROVAL_INVISIBLE_CHAR_SINGLE.test(cp) ? formatCodePointEscape(cp) : cp;
|
|
i += cp.length;
|
|
}
|
|
return truncateForDisplay(out);
|
|
}
|
|
|
|
function normalizePreview(commandText: string, commandPreview?: string | null): string | null {
|
|
const previewRaw = commandPreview?.trim() ?? "";
|
|
if (!previewRaw) {
|
|
return null;
|
|
}
|
|
const preview = sanitizeExecApprovalDisplayText(previewRaw);
|
|
if (preview === commandText) {
|
|
return null;
|
|
}
|
|
return preview;
|
|
}
|
|
|
|
export function resolveExecApprovalCommandDisplay(request: ExecApprovalRequestPayload): {
|
|
commandText: string;
|
|
commandPreview: string | null;
|
|
} {
|
|
const commandTextSource =
|
|
request.command ||
|
|
(request.host === "node" && request.systemRunPlan ? request.systemRunPlan.commandText : "");
|
|
const commandText = sanitizeExecApprovalDisplayText(commandTextSource);
|
|
const previewSource =
|
|
request.commandPreview ??
|
|
(request.host === "node" ? (request.systemRunPlan?.commandPreview ?? null) : null);
|
|
return {
|
|
commandText,
|
|
commandPreview: normalizePreview(commandText, previewSource),
|
|
};
|
|
}
|