mirror of
https://github.com/paperclipai/paperclip.git
synced 2026-11-05 03:51:54 +01:00
## Thinking Path > - Paperclip is the open source app people use to manage AI agents for work. > - Paperclip removes secrets from run logs, tool results and diagnostics with `redactSensitiveText` before it stores or shows them. > - The JSON field-name patterns in that function start with an unanchored `[A-Za-z0-9_-]*`. Each start position inside a long run of letters and digits scans to the end of the run, so the cost grows with the square of the run length. > - Base64 of text is one long run of letters and digits. A 40 KB base64 blob of source code in a shell command takes about 5 s to redact, and a large Gmail RAW tool result does not finish. > - The redaction runs on the main event loop, so one such chunk blocks every request and every run on the instance. > - This pull request makes these scans start only at the start of a run. The scans become linear, and the redacted output does not change. > - The benefit is that large base64 payloads no longer freeze the server, and every secret that was redacted before is still redacted. ## Linked Issues or Issue Description Fixes #15393. The tool content guard runs `redactSensitiveText` on tool results. A Gmail RAW body is base64url of a MIME message, which is one long letter/digit run. `validateToolContent` on a Gmail-shaped result with a 68 KB `raw` field takes 59 s on master and 19 ms with this change; a 5.4 MB `raw` field takes 0.65 s. Thanks @ts-beat for the clear report in #15393. Refs #12309 (false positives in the same patterns; this pull request does not change which text matches). I searched issues and pull requests for "redaction", "redactSensitiveText", "quadratic", "backtracking", "hangs the server" and "100% CPU". #15393 reports the hang. No open pull request fixes it. **Subsystem affected:** secret redaction (`server/src/redaction.ts`, `packages/adapter-utils/src/command-redaction.ts`), used by run logs, tool content guards and run diagnostics. **Problem:** `compactRunLogChunk` redacts each run log chunk before it truncates the chunk. A shell command such as `printf '%s' '<base64>' | base64 -d > file.ts` with base64 of source text makes one long letter/digit run. Base64 of ASCII text contains `+` or `/` only where every third byte is `>`, `?`, `~` or DEL, so random base64 does not show the problem. One `redactSensitiveText` call on such a tool-call line takes (local measurement): | Blob | Before | After | | --- | --- | --- | | 4 KB | 66 ms | 1.1 ms | | 8 KB | 272 ms | 0.5 ms | | 16 KB | 1.1 s | 1.0 ms | | 40 KB | 5.4 s | 2.5 ms | | 1 MB | not measured | 36 ms | The `\b`-anchored command patterns (`KEY=value`, JWT, `Authorization`) and the CLI option pattern have the same quadratic cost on dash-separated runs, for example base64url data. **Proposed solution:** start the unanchored field-name scans only where no name character precedes them (`(?<![A-Za-z0-9_-])`). A prefixed name such as `myApiKey`, `x_auth_token` or `GITHUB_TOKEN` still matches, because the scan from the start of the run absorbs the prefix. **Alternatives considered:** - Truncate or cap the input before redaction. Rejected: a cut can split a secret and leave part of it unredacted, and the scan stays quadratic below the cap. - Run the scan in a worker thread with a deadline, as #15393 suggests. That limits the damage, but the CPU cost stays, and a timeout must then drop or block the content. It is still useful as an additional defense. ## What Changed - `server/src/redaction.ts`: `JSON_SECRET_FIELD_TEXT_RE` and `ESCAPED_JSON_SECRET_FIELD_TEXT_RE` start a field name only where no name character precedes it. The lookbehind comes after the optional opening quote, so quoted keys behave as before. The `Authorization` word scan uses the same start instead of `\b`. - `packages/adapter-utils/src/command-redaction.ts`: the `KEY=value` pattern uses the same start. The JWT pattern starts only at the first word character of a run. A CLI option name can no longer span `--`, so each `--` starts its own short scan. - Redaction still runs before truncation. `compactRunLogChunk` does not change. Why the output does not change: when an old pattern could match from the middle of a run, it could also match from the start of that run, because the `[A-Za-z0-9_-]*` prefix absorbs the extra characters. The engine tries the start of the run first. So the old patterns never matched from the middle of a run, and the new start condition only skips positions that cannot produce a match. The old JWT pattern also always matched from the first word character of a run. A CLI option name always ends at the end of its run, so the match from the last `--` in the run ends at the same place. ## Screenshots No UI change. ## Verification - New tests: - prefixed secret field names in plain, single-quoted and escaped JSON are still redacted, also when the name is glued to the end of a long run; - a 48 KB base64-of-source tool call through `compactRunLogChunk` stays intact, the secret in the same command is redacted, and the call finishes within 1 s; - `redactSensitiveText` and `redactCommandText` handle 48-120 KB letter/digit, dash-separated, `--`-separated and base64url runs within 1 s, and redact the secret that follows each run; - `x--token`, `--x--token`, `a-b--api-key=`, `prefix-MY_API_KEY=` and `--<jwt>` are still redacted, and `my-token is here` is not changed. - Without the change, the timing tests fail: 9.3 s, 9.8 s, 32 s and 60 s against the 1 s budget. The correctness tests pass with and without the change. - Mutation check: I reverted each of the six pattern changes one at a time. Each revert makes at least one test fail. - Differential fuzzing: 1.1 million random inputs built from secret names, quotes, escaped quotes, dashes, dots, colons and JWT parts gave the same output from the old and new `redactSensitiveText`, `redactCommandText` and `redactDiagnosticText`. - `cd server && npx vitest run <files>`: all 33 server test files that use the redaction helpers, with embedded Postgres, 1730 passed. - `npx vitest run --project @paperclipai/adapter-utils packages/adapter-utils/src/command-redaction.test.ts`: 25 passed. - `npx vitest run --project @paperclipai/adapter-utils`, complete suite: 80 of 82 files passed in the parallel run. The 2 other files (`github-launcher-environment`, `sandbox-managed-runtime`) had 5 s timeouts in git-based tests and passed on rerun. ## Risks - Lookbehind assertions need ES2018. `redactCommandText` also runs in the UI bundle. Node and current browsers support lookbehind, and the server (`server/src/log-redaction.ts`) and the UI (`ui/src/components/task-chat/task-chat-attachments.ts`) already use it. - A change to a redaction pattern could leak a secret. The reasoning above, the unchanged existing tests and the differential fuzzing show the same output as before. - Not changed here, found while checking: `redactDiagnosticText` backtracks exponentially on an escaped JSON secret value made of many escaped backslashes with no closing quote, and `redactSensitiveText` redacts a value only up to its first quote of either kind (`{"apiKey":"abc\"tail"}` keeps `tail"`). Both behave the same before and after this pull request. ## Model Used - Claude Opus 5.5 (`claude-opus-5-5`) by Anthropic, in Claude Code, with tool use (shell, code search, and test runs) and extended thinking. A human directed the design choices. ## Checklist - [x] I have included a thinking path that traces from project context to this change - [x] I have specified the model used (with version and capability details) - [x] I have checked ROADMAP.md and confirmed this PR does not duplicate planned core work - [x] I have searched GitHub for duplicate or related PRs and linked them above - [x] I have either (a) linked existing issues with `Fixes: #` / `Closes #` / `Refs #` OR (b) described the issue in-PR following the relevant issue template - [x] I have not referenced internal/instance-local Paperclip issues or links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip` URLs) - [x] My branch name describes the change (e.g. `docs/...`, `fix/...`) and contains no internal Paperclip ticket id or instance-derived details - [x] I have run tests locally and they pass - [x] I have added or updated tests where applicable - [ ] I have updated relevant documentation to reflect my changes - [x] I have considered and documented any risks above - [x] All Paperclip CI gates are green - [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups - [x] I will address all Greptile and reviewer comments before requesting merge
150 lines
5.6 KiB
TypeScript
150 lines
5.6 KiB
TypeScript
export const REDACTED_COMMAND_TEXT_VALUE = "***REDACTED***";
|
|
|
|
// Public Executor helper addresses retained for callers of this predicate.
|
|
// Dotted addresses alone do not meet the JWT credential heuristic.
|
|
const PUBLIC_EXECUTOR_TOOL_SELECTORS = new Set([
|
|
"executor.coreTools.integrations.list",
|
|
"executor.coreTools.connections.list",
|
|
"executor.coreTools.policies.list",
|
|
]);
|
|
export function isPublicExecutorToolSelector(value: string): boolean {
|
|
return PUBLIC_EXECUTOR_TOOL_SELECTORS.has(value);
|
|
}
|
|
|
|
const SECRET_NAME_WORDS = String.raw`(?:api[-_]?key|(?:access[-_]?|auth[-_]?)?token|token|authorization(?:[-_]?code)?|bearer|secrets?|passwd|passwords?|credentials?|jwt|private[-_]?key|cookie|connectionstring)(?:[-_]?(?:value|header|prod(?:uction)?|dev(?:elopment)?|test|staging|primary|secondary))*`;
|
|
const SECRET_NAME_PATTERN = String.raw`[A-Za-z0-9_-]*${SECRET_NAME_WORDS}`;
|
|
const SECRET_NAME_START = String.raw`(?<![A-Za-z0-9_-])`;
|
|
const SECRET_OPTION_NAME_PATTERN = String.raw`(?:[A-Za-z0-9_]|-(?!-))*${SECRET_NAME_WORDS}`;
|
|
|
|
const COMMAND_CLI_SECRET_OPTION_RE = new RegExp(
|
|
String.raw`(\B-{1,2}${SECRET_OPTION_NAME_PATTERN}(?:\s+|=)(["']?))[^\s"'` +
|
|
"`" +
|
|
String.raw`]+(\2)`,
|
|
"gi",
|
|
);
|
|
const COMMAND_ENV_SECRET_ASSIGNMENT_RE = new RegExp(
|
|
String.raw`(${SECRET_NAME_START}${SECRET_NAME_PATTERN}\s*=\s*)(?:(\\["'])([\s\S]*?)\2|(["'])([^"'` +
|
|
"`" +
|
|
String.raw`\r\n]*)\4|([^\s"'` +
|
|
"`" +
|
|
String.raw`]+))`,
|
|
"gi",
|
|
);
|
|
const COMMAND_AUTHORIZATION_BEARER_RE =
|
|
/(\bAuthorization\s*:\s*Bearer\s+)[^\s"'`]+/gi;
|
|
const COMMAND_OPENAI_KEY_RE = /\bsk-[A-Za-z0-9_-]{12,}\b/g;
|
|
const COMMAND_GITHUB_TOKEN_RE = /\bgh[pousr]_[A-Za-z0-9_]{20,}\b/g;
|
|
const COMMAND_JWT_RE =
|
|
/\b(?<![A-Za-z0-9_]-*)[A-Za-z0-9_-]{8,}(?:\.[A-Za-z0-9_-]{8,}){2}(?:\.[A-Za-z0-9_-]{8,}\.[A-Za-z0-9_-]{8,})?\b/g;
|
|
/** Recognize encoded JSON headers, without treating dotted identifiers as tokens. */
|
|
export function looksLikeCredentialJwt(value: string): boolean {
|
|
const segments = value.split(".");
|
|
if (![3, 5].includes(segments.length)) return false;
|
|
if (segments.some((part, index) => {
|
|
// Compact JWE direct encryption has no encrypted-key value.
|
|
if (segments.length === 5 && index === 1 && part === "") return false;
|
|
return !/^[A-Za-z0-9_-]+$/.test(part);
|
|
})) return false;
|
|
if (segments[0].startsWith("eyJ")) return true;
|
|
try {
|
|
const header = JSON.parse(atob(segments[0].replace(/-/g, "+").replace(/_/g, "/")));
|
|
return typeof header === "object" && header !== null && typeof header.alg === "string";
|
|
} catch {
|
|
return false;
|
|
}
|
|
}
|
|
|
|
const COMMAND_SECRET_HINTS = [
|
|
"api",
|
|
"key",
|
|
"token",
|
|
"auth",
|
|
"bearer",
|
|
"secret",
|
|
"pass",
|
|
"credential",
|
|
"jwt",
|
|
"private",
|
|
"cookie",
|
|
"connectionstring",
|
|
"sk-",
|
|
"ghp_",
|
|
"gho_",
|
|
"ghu_",
|
|
"ghs_",
|
|
"ghr_",
|
|
] as const;
|
|
|
|
function maybeContainsSecretText(command: string) {
|
|
const lower = command.toLowerCase();
|
|
return (
|
|
COMMAND_SECRET_HINTS.some((hint) => lower.includes(hint)) ||
|
|
command.includes(".")
|
|
);
|
|
}
|
|
|
|
export function redactCommandText(
|
|
command: string,
|
|
redactedValue = REDACTED_COMMAND_TEXT_VALUE,
|
|
): string {
|
|
if (!maybeContainsSecretText(command)) return command;
|
|
return command
|
|
.replace(COMMAND_AUTHORIZATION_BEARER_RE, `$1${redactedValue}`)
|
|
.replace(COMMAND_CLI_SECRET_OPTION_RE, `$1${redactedValue}$3`)
|
|
.replace(
|
|
COMMAND_ENV_SECRET_ASSIGNMENT_RE,
|
|
(
|
|
_match,
|
|
prefix: string,
|
|
escapedQuote: string | undefined,
|
|
_escapedValue: string | undefined,
|
|
rawQuote: string | undefined,
|
|
) => {
|
|
const quote = escapedQuote ?? rawQuote;
|
|
return quote
|
|
? `${prefix}${quote}${redactedValue}${quote}`
|
|
: `${prefix}${redactedValue}`;
|
|
},
|
|
)
|
|
.replace(COMMAND_OPENAI_KEY_RE, redactedValue)
|
|
.replace(COMMAND_GITHUB_TOKEN_RE, redactedValue)
|
|
.replace(COMMAND_JWT_RE, (match) => looksLikeCredentialJwt(match) ? redactedValue : match);
|
|
}
|
|
|
|
// A JSON secret field is a key/value pair such as `"token":"opaque-value"`. The
|
|
// command redaction handles shell `KEY=value` syntax only. A sandbox diagnostic
|
|
// can also carry a serialized JSON error, so the sanitizer must redact the JSON
|
|
// form too. The value body consumes JSON escape sequences. An escaped quote
|
|
// (`\"`) inside the value does not end the match early.
|
|
const JSON_SECRET_FIELD_RE = new RegExp(
|
|
String.raw`("(?:${SECRET_NAME_PATTERN})"\s*:\s*")(?:\\[\s\S]|[^"\\])*(")`,
|
|
"gi",
|
|
);
|
|
// An escaped JSON secret field is the same pair inside a JSON string. The double
|
|
// quote appears as `\"` and a backslash appears as `\\`. The value body
|
|
// consumes the doubled escape sequences. An escaped quote inside the value does
|
|
// not end the match early. The value ends at the next unescaped `\"`.
|
|
const JSON_ESCAPED_SECRET_FIELD_RE = new RegExp(
|
|
String.raw`(\\"(?:${SECRET_NAME_PATTERN})\\"\s*:\s*\\")(?:\\\\\\\\|\\\\\\"|\\\\[\s\S]|[^\\"])*(\\")`,
|
|
"gi",
|
|
);
|
|
|
|
/**
|
|
* Redact secrets from an untrusted diagnostic string.
|
|
*
|
|
* The function first runs the command redaction. The command redaction handles
|
|
* shell `KEY=value` assignments, CLI secret options, bearer headers, and common
|
|
* token shapes. The function then redacts JSON and escaped-JSON secret fields,
|
|
* because a sandbox diagnostic can carry a serialized JSON error such as
|
|
* `{"token":"opaque-value"}`. The caller must still bound the length after this
|
|
* step.
|
|
*/
|
|
export function redactDiagnosticText(
|
|
text: string,
|
|
redactedValue = REDACTED_COMMAND_TEXT_VALUE,
|
|
): string {
|
|
return redactCommandText(text, redactedValue)
|
|
.replace(JSON_ESCAPED_SECRET_FIELD_RE, `$1${redactedValue}$2`)
|
|
.replace(JSON_SECRET_FIELD_RE, `$1${redactedValue}$2`);
|
|
}
|