{"title":"Everything That Broke The Week Of November 16th","html_url":"https://deallocatedobjects.com/posts/everything-that-broke-the-week-of-november-16th","source_url":"","markdown_content":"A month ago I wrote up [everything that broke on October 21st](/posts/everything-that-broke-on-october-21st), a day when unrelated security disasters kept landing one after another. This week did the same thing, only spread across five days. nginx, Java and Ruby each got a pre-auth remote code execution bug within about 48 hours of each other, PHP shipped its second big batch of fixes in a month, and a bunch of the internet's biggest websites spent most of Monday serving a card skimmer.\n\nHere's what happened, roughly ordered by how much it should worry you.\n\n### nginx2root\n\nThe big one. On Wednesday F5 shipped [nginx **1.31.9** and **1.30.7**](https://blog.qualys.com/product-tech/patch-management/2026/11/18/nginx2root-five-nginx-flaws-led-by-cve-2026-123007-how-to-find-mitigate-and-patch-now) with fixes for ten vulnerabilities. Five of them are unauthenticated memory-corruption bugs that [Xint](https://xint.io/blog/nginx2root-remote-code-execution) grouped together as \"nginx2root.\" The Hacker News thread for it hit 1,508 points.\n\nThe lead bug, **CVE-2026-123007** (CVSS v4 9.2), has been in nginx since HTTP/2 support arrived in **1.9.5**, back in 2015. Every official release from then through 1.31.8 has it. It's an integer overflow in HTTP/2 flow control. When a client cancels a stream before nginx sends its queued DATA frames, `ngx_http_v2_filter_cleanup()` gives the reserved credit back to the connection's send window without checking it against the protocol maximum. A client can return the same credit twice (once with a `WINDOW_UPDATE`, once through the cleanup) and push the window past 2³¹−1.\n\nFrom there it gets fun. The check in `ngx_http_v2_state_window_update()` computes the remaining room as `NGX_HTTP_V2_MAX_WINDOW - h2c->send_window` in an unsigned `size_t`, which wraps to a huge number once the window is already over the limit, so every further increment is accepted. The window then gets cast to a signed `off_t`, goes negative, and ends up as the length of an `ngx_memcpy` in the TLS code, about 2⁶⁴ bytes. That would normally just crash. But glibc's x86-64 `memcpy` wraps its end pointer around below the destination and runs its backward copy loop only once, so the copy touches a small region instead. Xint turned that into an out-of-bounds write into TLS allocation metadata, a leak of nginx's heap and image addresses back through the TLS response, and control of a cleanup function pointer that runs a command as the worker. They report it working against public HTTPS endpoints on AWS, GCP and Azure, under normal traffic.\n\nThe other four:\n\n- **CVE-2026-123633**: a use-after-free when nginx runs low on connection slots and drains idle connections. An HTTP/2 connection with a half-read frame can stay on the reusable list, get closed by its own stream's upstream connect, and leave the proxy code working on a freed request. Needs HTTP/2 plus `proxy_pass`. Since 1.19.7.\n- **CVE-2026-121810**: the same idea in HTTP/3. Since 1.25.0.\n- **CVE-2026-123634**: a use-after-free in the DNS resolver, reachable with a variable in `proxy_pass`, a `resolver` and a subrequest like `mirror` or SSI. The attacker needs to control a DNS server nginx will ask. Since 1.13.4.\n- **CVE-2026-113425**: a use-after-free in the mail proxy's `auth_http` path. Since **0.5.15**, which is from 2007.\n\nXint found the HTTP/2 bug by hand while getting ready for ZeroDay Cloud 2026, then a fully automated scan found the other four. They also found an unpatched Linux kernel SCTP use-after-free (in the kernel since v4.15, no CVE yet) that takes you from the nginx worker to root, and from inside a container to the host. Exploit code for CVE-2026-123007 is public, though Qualys describes it as slow and fiddly: minutes of HTTP/2 control frames over one long-lived connection. No exploitation in the wild reported so far.\n\nA few things worth knowing when you patch:\n\n- **Restart, don't reload.** A reload re-reads the config but keeps running the old executable. Old workers also hang around serving existing connections until those close.\n- **Your version number might not change.** Distros backport the fix. On Ubuntu 24.04 the fixed package is `1.24.0-2ubuntu7.20`, and `nginx -v` still says `1.24.0`. Check `dpkg -l nginx` (or `apt changelog nginx`) instead.\n- **ingress-nginx is done.** The community Kubernetes controller was archived in March and won't get this fix. F5's own NGINX Ingress Controller 5.8.0 has it.\n- **The log is quiet.** `WINDOW_UPDATE` and `RST_STREAM` frames aren't requests, so the access log shows nothing. Repeated worker crashes with signal 11 are the better tell.\n\nIf you can't upgrade yet, turning off HTTP/2 blocks the two HTTP/2 bugs:\n\n```nginx\nhttp2 off;   # 1.25.1 and newer. Older: remove \"http2\" from the listen line\n```\n\nAnd if you want to keep SCTP from loading until there's a kernel fix (this only blocks loading, it won't unload a module that's already in):\n\n```bash\necho \"install sctp /bin/false\" | sudo tee /etc/modprobe.d/disable-sctp.conf\n```\n\n### Java4Shell\n\nOn Tuesday [Calif](https://calif.io/java4shell) disclosed **CVE-2026-114938**, a heap overflow in the JVM's Base64 decoder. Specifically, in HotSpot's hand-written x86-64 assembly for `Base64.Decoder`, the AVX-512 VBMI version the JIT swaps in once a decode call gets hot.\n\nWhen the input ends in padding, that code computes the output length for the last chunk as zero and then subtracts one for each `=`, landing on −1. The mask it builds from that length comes out as all ones, so the masked 64-byte store writes every byte. Calif's trigger is 64 `A`s and one `=`. That decodes into a 50-byte array, the store starts at offset 48, and 62 bytes land past the end of it. Sixty-five characters of Base64.\n\nIt affects OpenJDK **17.0.6 and newer**, plus 21, 25 and 27, and every downstream build (Temurin, Corretto, Zulu, Microsoft's, Red Hat's and so on). Java 8 and 11 don't have the AVX-512 decoder. It also only matters on CPUs with AVX-512 VBMI: Intel Ice Lake Xeons and newer, AMD Zen 4 and newer. ARM is fine, so anything on Graviton or Apple silicon is off the hook. To check a Linux box:\n\n```bash\ngrep -m1 '^flags' /proc/cpuinfo | tr ' ' '\\n' | grep -E '^(avx512f|avx512bw|avx512vbmi|bmi2)$'\n```\n\nAll four in the output means the vulnerable path can run.\n\nCalif says it reached pre-auth code execution in Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate, through ordinary things like authentication and file uploads (think Basic auth headers and JWTs). They're holding back the product-specific chains for now. samczsun also showed a client-side version, where a Java program that connects to a malicious TLS server gets owned.\n\nThe backstory is the wild part. Quang Luong at Calif found it by pointing OpenAI's Codex at the JIT's intrinsics: an hour to find, two more to write a working exploit, under $200 in tokens. Then they went looking and found it was **already public**. Someone had reported the crash in an Adoptium issue on June 9th, posted a proof of concept on September 8th, and another developer root-caused it the same day and opened a fix PR that got closed without merging. It sat there for months. Calif reported it to Oracle on October 24th, Oracle fixed it on November 13th, and the patches shipped Tuesday.\n\n[OX Security](https://www.ox.security/blog/java4shell-cve-2026-114938-rce-risk/) pushed back on the panic a bit. Their estimate is that about 0.8% of JVMs can be crashed remotely without logging in, and about 0.12% are realistic code execution targets, because each exploit has to be tuned to one app's heap. Fair, but crashing your SSO server with a 65-character string is bad enough.\n\nPatch to the November 17th builds. If you can't yet, there's a flag that turns off the intrinsic, and it needs a JVM restart to take effect:\n\n```bash\n-XX:-UseBASE64Intrinsics\n```\n\n`-XX:UseAVX=2` also works on x86, but don't put it in a config that's shared with ARM machines, because an ARM JVM won't start with it.\n\n### Ruby2ßhell\n\nWednesday was Ruby's turn. Ruby **3.3.14**, **3.4.13** and **4.0.10** fix a heap overflow in string case conversion that [Veria Labs](https://equixly.com/blog/2026/11/19/ruby2shell-pre-auth-heap-overflow-in-ruby/) named Ruby2ßhell, and yes, that's an actual ß in the name.\n\nThe German ß has no single-character uppercase form, so upcasing it produces `SS`. In 14 old single-byte encodings like ISO-8859-1 and Windows-1252, the case conversion loop checks that there's room for one more byte, then writes two when it hits a ß. The output buffer gets 20 bytes of slack, so 21 or more ß's overflow it. It's reachable through `upcase`, `swapcase`, `downcase(:fold)` and `casecmp?`, and every Ruby back to **2.4.0** has it.\n\nVeria says a plain `rails new` app is reachable before login, but they haven't said which input yet, and they plan to publish their exploit around November 25th. Ruby's advisory says the team is \"aware of PoCs that prove this flaw can be used to trigger RCE in some scenarios.\" It credits Stephan Lipp of Ada Logics and Anthropic with the find.\n\nThe fix is in the interpreter, so bumping Rails or any gem does nothing. Ruby 3.2 and older are end of life and won't get a fix. And a Puma phased restart keeps the old Ruby running, so do a full restart. If you self-host GitLab or Mastodon, you need their updated packages.\n\n### PHP, Round Two\n\nPHP shipped **8.5.13**, **8.4.28**, **8.3.37** and **8.2.36** on Thursday (along with 8.6.0), covering about 17 advisories, seven of them rated critical. It's the follow-up to last month's PHP2Shell, and the session upload progress feature shows up again.\n\n**CVE-2026-115560** is the third RCE PHP reserved back in October. Upload progress data gets written into the session under one `session.serialize_handler` and read back with another, which hands attacker bytes to `unserialize`. It needs a non-default setup, like two apps sharing a session store with different handlers. The fix is the part to notice: **`session.upload_progress.enabled` is now off by default**, so if something of yours shows upload progress bars from the session, they'll stop working after the upgrade until you turn it back on.\n\nThe other one to read about is **CVE-2026-129289**, request smuggling in PHP-FPM. Once a script opens `php://input`, PHP stops draining any body bytes the script doesn't read. If the web server reuses FastCGI connections, those leftover bytes get parsed as the next FastCGI request, and a forged one can set `allow_url_include=1` and `auto_prepend_file=php://input` to run the body as PHP. Connection reuse isn't the default, so check for it:\n\n```bash\ngrep -rn \"fastcgi_keep_conn\" /etc/nginx/\n```\n\nIf it says `on`, turn it off until you've upgraded.\n\nPHP2Shell itself is still causing damage. [BleepingComputer reported](https://www.bleepingcomputer.com/news/security/check-point-gateways-hacked-via-php2shell-to-deploy-vmware-esxi-wiper/) attackers getting into unpatched Check Point gateways through their PHP-based Mobile Access portal, using a second bug (CVE-2026-123039) for root, then moving to VMware ESXi hosts and running a wiper. If you run those gateways, look for session files containing `upload_progress_pwn`.\n\n### Optimizely\n\nEarly Monday someone injected JavaScript into Optimizely Web Experimentation projects, and the A/B testing scripts those sites load from Optimizely's CDN started serving a crypto address swapper and a card skimmer. [Report URI](https://www.securityweek.com/optimizely-vulnerability-exploited-to-serve-malware-on-major-websites/) checked 1,611 Optimizely scripts and found 188 infected, belonging to 178 organizations, including the BBC, CNN, Zoom, GitLab, Salesforce, HSBC and IKEA.\n\nThe first injection they saw was at 04:21 UTC on the 16th, and everything was clean by 23:11 UTC. Optimizely says the attacker got in through a flaw in its experimentation account signup flow, then wrote a loader into each project's custom JavaScript field. StepSecurity found that card numbers typed into third-party payment iframes were out of the skimmer's reach, which is a good argument for those iframes.\n\nThis is the problem with any script tag pointing at someone else's server. Your page runs whatever is there today. A strict Content Security Policy won't stop a script you've already allowed, but it can stop that script from sending data to an attacker's domain with `connect-src`.\n\n### Barracuda And ARTEX Again\n\nRemember ARTEX, the open-source agent pentest framework Microsoft tied to the ConnectWise attack last month? [Barracuda says](https://www.bleepingcomputer.com/news/security/barracuda-links-cloud-breach-to-artex-ai-powered-cyberattack/) it showed up in their cloud too. The way in was a cloud access key Barracuda had given an outside vendor for a past project and never retired. It was sitting in a config file in the vendor's GitHub repo.\n\nOver 44 hours on November 1st and 2nd the attacker worked through about 10 Barracuda cloud accounts, reached source code, Kubernetes clusters and the system storing customer credentials for Barracuda's XDR integrations, tried to launch a container that mounted the host's file system, and tried to make itself a cloud admin (that part was blocked). Barracuda is asking affected XDR customers to recreate those credentials.\n\nGo look at your old access keys. The ones from finished projects are the dangerous ones, because nobody's watching them.\n\n### Qubole And Expedia\n\nSame shape, bigger blast radius. [Wiz found](https://www.wiz.io/blog/qubole-cross-account-access-abused-in-the-wild) that someone with AWS credentials belonging to Qubole, a big-data platform, spent October 14th through 26th assuming the integration roles Qubole customers had created in their own AWS accounts. Every request had the right External ID, so it all looked legitimate. Many of those roles had been sitting unused for a long time. Sentra counted about 100 secrets read from Secrets Manager, Parameter Store and Lambda environment variables in seven minutes.\n\nExpedia confirmed attackers reached it through \"a compromised third-party data platform provider,\" and a group calling itself n0n claims more than 830 million loyalty records. Expedia hasn't confirmed that number.\n\nIf you ever connected Qubole to AWS, search CloudTrail from October 1st for role sessions starting with `quiam-`, and do it soon. CloudTrail's event history only keeps 90 days, so the October 14th activity starts disappearing in mid-January.\n\n### SCKit\n\nThis one started last Wednesday, but its second wave hit on the 13th. A third-party ScreenConnect extension called RMM+ (versions 1.0.31 and 1.0.32) accepted a blank shared key, and that key was blank by default. [Huntress watched](https://www.huntress.com/blog/sckit-worm-rogue-screenconnect) an attacker use it to skip ScreenConnect's accounts and MFA entirely and push its own ScreenConnect client as SYSTEM to 288 machines at 48 companies in about 90 minutes, then 205 more machines at 41 companies two days later. ConnectWise disabled five extensions across roughly 400 partner environments. On-prem installs have to disable them by hand.\n\n### The Rest Of The Week\n\n- CISA added a LiteLLM sandbox escape (CVE-2026-40217) to its Known Exploited Vulnerabilities list. The `/guardrails/test_custom_code` endpoint runs user Python in a homemade sandbox, and the default Docker image runs the proxy as root. It needs an admin key. Fixed in 1.83.11.\n- uBlock Origin fixed a universal XSS where a malicious site could trick the element picker into saving a trusted filter that ran JavaScript on other sites. Update to 1.75.4 or newer.\n- OpenSSL's CVE-2026-121481 showed up with a scary CVSS 9.8 on the CVE record. OpenSSL itself rates it Low: it needs a Base64 input of about 1.6 GB, and nobody has found a way for an attacker to supply one. Fixes land December 1st in 4.0.4 and 4.1.1.\n- Infoblox confirmed attackers got at company information through a third-party vendor.\n- The Gentlemen ransomware group added NVIDIA to its leak site with no proof. NVIDIA says it's looking into an incident at a third-party partner.\n\nGo restart your stuff, y'all. A reload won't pick up most of this week's fixes.","html_content":"<p>A month ago I wrote up <a href=\"https://deallocatedobjects.com/posts/everything-that-broke-on-october-21st\">everything that broke on October 21st</a>, a day when unrelated security disasters kept landing one after another. This week did the same thing, only spread across five days. nginx, Java and Ruby each got a pre-auth remote code execution bug within about 48 hours of each other, PHP shipped its second big batch of fixes in a month, and a bunch of the internet’s biggest websites spent most of Monday serving a card skimmer.</p>\n<p>Here’s what happened, roughly ordered by how much it should worry you.</p>\n<h2 id=\"nginx2root\" data-level=\"3\">nginx2root</h2>\n<p>The big one. On Wednesday F5 shipped <a href=\"https://blog.qualys.com/product-tech/patch-management/2026/11/18/nginx2root-five-nginx-flaws-led-by-cve-2026-123007-how-to-find-mitigate-and-patch-now\">nginx <strong>1.31.9</strong> and <strong>1.30.7</strong></a> with fixes for ten vulnerabilities. Five of them are unauthenticated memory-corruption bugs that <a href=\"https://xint.io/blog/nginx2root-remote-code-execution\">Xint</a> grouped together as “nginx2root.” The Hacker News thread for it hit 1,508 points.</p>\n<p>The lead bug, <strong>CVE-2026-123007</strong> (CVSS v4 9.2), has been in nginx since HTTP/2 support arrived in <strong>1.9.5</strong>, back in 2015. Every official release from then through 1.31.8 has it. It’s an integer overflow in HTTP/2 flow control. When a client cancels a stream before nginx sends its queued DATA frames, <code>ngx_http_v2_filter_cleanup()</code> gives the reserved credit back to the connection’s send window without checking it against the protocol maximum. A client can return the same credit twice (once with a <code>WINDOW_UPDATE</code>, once through the cleanup) and push the window past 2³¹−1.</p>\n<p>From there it gets fun. The check in <code>ngx_http_v2_state_window_update()</code> computes the remaining room as <code>NGX_HTTP_V2_MAX_WINDOW - h2c-&gt;send_window</code> in an unsigned <code>size_t</code>, which wraps to a huge number once the window is already over the limit, so every further increment is accepted. The window then gets cast to a signed <code>off_t</code>, goes negative, and ends up as the length of an <code>ngx_memcpy</code> in the TLS code, about 2⁶⁴ bytes. That would normally just crash. But glibc’s x86-64 <code>memcpy</code> wraps its end pointer around below the destination and runs its backward copy loop only once, so the copy touches a small region instead. Xint turned that into an out-of-bounds write into TLS allocation metadata, a leak of nginx’s heap and image addresses back through the TLS response, and control of a cleanup function pointer that runs a command as the worker. They report it working against public HTTPS endpoints on AWS, GCP and Azure, under normal traffic.</p>\n<p>The other four:</p>\n<ul>\n<li><strong>CVE-2026-123633</strong>: a use-after-free when nginx runs low on connection slots and drains idle connections. An HTTP/2 connection with a half-read frame can stay on the reusable list, get closed by its own stream’s upstream connect, and leave the proxy code working on a freed request. Needs HTTP/2 plus <code>proxy_pass</code>. Since 1.19.7.</li>\n<li><strong>CVE-2026-121810</strong>: the same idea in HTTP/3. Since 1.25.0.</li>\n<li><strong>CVE-2026-123634</strong>: a use-after-free in the DNS resolver, reachable with a variable in <code>proxy_pass</code>, a <code>resolver</code> and a subrequest like <code>mirror</code> or SSI. The attacker needs to control a DNS server nginx will ask. Since 1.13.4.</li>\n<li><strong>CVE-2026-113425</strong>: a use-after-free in the mail proxy’s <code>auth_http</code> path. Since <strong>0.5.15</strong>, which is from 2007.</li>\n</ul>\n<p>Xint found the HTTP/2 bug by hand while getting ready for ZeroDay Cloud 2026, then a fully automated scan found the other four. They also found an unpatched Linux kernel SCTP use-after-free (in the kernel since v4.15, no CVE yet) that takes you from the nginx worker to root, and from inside a container to the host. Exploit code for CVE-2026-123007 is public, though Qualys describes it as slow and fiddly: minutes of HTTP/2 control frames over one long-lived connection. No exploitation in the wild reported so far.</p>\n<p>A few things worth knowing when you patch:</p>\n<ul>\n<li><strong>Restart, don’t reload.</strong> A reload re-reads the config but keeps running the old executable. Old workers also hang around serving existing connections until those close.</li>\n<li><strong>Your version number might not change.</strong> Distros backport the fix. On Ubuntu 24.04 the fixed package is <code>1.24.0-2ubuntu7.20</code>, and <code>nginx -v</code> still says <code>1.24.0</code>. Check <code>dpkg -l nginx</code> (or <code>apt changelog nginx</code>) instead.</li>\n<li><strong>ingress-nginx is done.</strong> The community Kubernetes controller was archived in March and won’t get this fix. F5’s own NGINX Ingress Controller 5.8.0 has it.</li>\n<li><strong>The log is quiet.</strong> <code>WINDOW_UPDATE</code> and <code>RST_STREAM</code> frames aren’t requests, so the access log shows nothing. Repeated worker crashes with signal 11 are the better tell.</li>\n</ul>\n<p>If you can’t upgrade yet, turning off HTTP/2 blocks the two HTTP/2 bugs:</p>\n<figure class=\"code\" data-lang=\"nginx\"><figcaption><span class=\"lang\">nginx</span></figcaption><pre class=\"shiki shiki-themes github-light github-dark\" style=\"--shiki-light:#24292e;--shiki-dark:#e1e4e8;--shiki-light-bg:#fff;--shiki-dark-bg:#24292e\" tabindex=\"0\"><code><span class=\"line\"><span style=\"--shiki-light:#D73A49;--shiki-dark:#F97583\">http2 </span><span style=\"--shiki-light:#005CC5;--shiki-dark:#79B8FF\">off</span><span style=\"--shiki-light:#24292E;--shiki-dark:#E1E4E8\">;   </span><span style=\"--shiki-light:#6A737D;--shiki-dark:#6A737D\"># 1.25.1 and newer. Older: remove \"http2\" from the listen line</span></span></code></pre></figure><p>And if you want to keep SCTP from loading until there’s a kernel fix (this only blocks loading, it won’t unload a module that’s already in):</p>\n<figure class=\"code\" data-lang=\"bash\"><figcaption><span class=\"lang\">Shell</span></figcaption><pre class=\"shiki shiki-themes github-light github-dark\" style=\"--shiki-light:#24292e;--shiki-dark:#e1e4e8;--shiki-light-bg:#fff;--shiki-dark-bg:#24292e\" tabindex=\"0\"><code><span class=\"line\"><span style=\"--shiki-light:#005CC5;--shiki-dark:#79B8FF\">echo</span><span style=\"--shiki-light:#032F62;--shiki-dark:#9ECBFF\"> \"install sctp /bin/false\"</span><span style=\"--shiki-light:#D73A49;--shiki-dark:#F97583\"> |</span><span style=\"--shiki-light:#6F42C1;--shiki-dark:#B392F0\"> sudo</span><span style=\"--shiki-light:#032F62;--shiki-dark:#9ECBFF\"> tee</span><span style=\"--shiki-light:#032F62;--shiki-dark:#9ECBFF\"> /etc/modprobe.d/disable-sctp.conf</span></span></code></pre></figure><h2 id=\"java4shell\" data-level=\"3\">Java4Shell</h2>\n<p>On Tuesday <a href=\"https://calif.io/java4shell\">Calif</a> disclosed <strong>CVE-2026-114938</strong>, a heap overflow in the JVM’s Base64 decoder. Specifically, in HotSpot’s hand-written x86-64 assembly for <code>Base64.Decoder</code>, the AVX-512 VBMI version the JIT swaps in once a decode call gets hot.</p>\n<p>When the input ends in padding, that code computes the output length for the last chunk as zero and then subtracts one for each <code>=</code>, landing on −1. The mask it builds from that length comes out as all ones, so the masked 64-byte store writes every byte. Calif’s trigger is 64 <code>A</code>s and one <code>=</code>. That decodes into a 50-byte array, the store starts at offset 48, and 62 bytes land past the end of it. Sixty-five characters of Base64.</p>\n<p>It affects OpenJDK <strong>17.0.6 and newer</strong>, plus 21, 25 and 27, and every downstream build (Temurin, Corretto, Zulu, Microsoft’s, Red Hat’s and so on). Java 8 and 11 don’t have the AVX-512 decoder. It also only matters on CPUs with AVX-512 VBMI: Intel Ice Lake Xeons and newer, AMD Zen 4 and newer. ARM is fine, so anything on Graviton or Apple silicon is off the hook. To check a Linux box:</p>\n<figure class=\"code\" data-lang=\"bash\"><figcaption><span class=\"lang\">Shell</span></figcaption><pre class=\"shiki shiki-themes github-light github-dark\" style=\"--shiki-light:#24292e;--shiki-dark:#e1e4e8;--shiki-light-bg:#fff;--shiki-dark-bg:#24292e\" tabindex=\"0\"><code><span class=\"line\"><span style=\"--shiki-light:#6F42C1;--shiki-dark:#B392F0\">grep</span><span style=\"--shiki-light:#005CC5;--shiki-dark:#79B8FF\"> -m1</span><span style=\"--shiki-light:#032F62;--shiki-dark:#9ECBFF\"> '^flags'</span><span style=\"--shiki-light:#032F62;--shiki-dark:#9ECBFF\"> /proc/cpuinfo</span><span style=\"--shiki-light:#D73A49;--shiki-dark:#F97583\"> |</span><span style=\"--shiki-light:#6F42C1;--shiki-dark:#B392F0\"> tr</span><span style=\"--shiki-light:#032F62;--shiki-dark:#9ECBFF\"> ' '</span><span style=\"--shiki-light:#032F62;--shiki-dark:#9ECBFF\"> '\\n'</span><span style=\"--shiki-light:#D73A49;--shiki-dark:#F97583\"> |</span><span style=\"--shiki-light:#6F42C1;--shiki-dark:#B392F0\"> grep</span><span style=\"--shiki-light:#005CC5;--shiki-dark:#79B8FF\"> -E</span><span style=\"--shiki-light:#032F62;--shiki-dark:#9ECBFF\"> '^(avx512f|avx512bw|avx512vbmi|bmi2)$'</span></span></code></pre></figure><p>All four in the output means the vulnerable path can run.</p>\n<p>Calif says it reached pre-auth code execution in Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate, through ordinary things like authentication and file uploads (think Basic auth headers and JWTs). They’re holding back the product-specific chains for now. samczsun also showed a client-side version, where a Java program that connects to a malicious TLS server gets owned.</p>\n<p>The backstory is the wild part. Quang Luong at Calif found it by pointing OpenAI’s Codex at the JIT’s intrinsics: an hour to find, two more to write a working exploit, under $200 in tokens. Then they went looking and found it was <strong>already public</strong>. Someone had reported the crash in an Adoptium issue on June 9th, posted a proof of concept on September 8th, and another developer root-caused it the same day and opened a fix PR that got closed without merging. It sat there for months. Calif reported it to Oracle on October 24th, Oracle fixed it on November 13th, and the patches shipped Tuesday.</p>\n<p><a href=\"https://www.ox.security/blog/java4shell-cve-2026-114938-rce-risk/\">OX Security</a> pushed back on the panic a bit. Their estimate is that about 0.8% of JVMs can be crashed remotely without logging in, and about 0.12% are realistic code execution targets, because each exploit has to be tuned to one app’s heap. Fair, but crashing your SSO server with a 65-character string is bad enough.</p>\n<p>Patch to the November 17th builds. If you can’t yet, there’s a flag that turns off the intrinsic, and it needs a JVM restart to take effect:</p>\n<figure class=\"code\" data-lang=\"bash\"><figcaption><span class=\"lang\">Shell</span></figcaption><pre class=\"shiki shiki-themes github-light github-dark\" style=\"--shiki-light:#24292e;--shiki-dark:#e1e4e8;--shiki-light-bg:#fff;--shiki-dark-bg:#24292e\" tabindex=\"0\"><code><span class=\"line\"><span style=\"--shiki-light:#6F42C1;--shiki-dark:#B392F0\">-XX:-UseBASE64Intrinsics</span></span></code></pre></figure><p><code>-XX:UseAVX=2</code> also works on x86, but don’t put it in a config that’s shared with ARM machines, because an ARM JVM won’t start with it.</p>\n<h2 id=\"ruby2-hell\" data-level=\"3\">Ruby2ßhell</h2>\n<p>Wednesday was Ruby’s turn. Ruby <strong>3.3.14</strong>, <strong>3.4.13</strong> and <strong>4.0.10</strong> fix a heap overflow in string case conversion that <a href=\"https://equixly.com/blog/2026/11/19/ruby2shell-pre-auth-heap-overflow-in-ruby/\">Veria Labs</a> named Ruby2ßhell, and yes, that’s an actual ß in the name.</p>\n<p>The German ß has no single-character uppercase form, so upcasing it produces <code>SS</code>. In 14 old single-byte encodings like ISO-8859-1 and Windows-1252, the case conversion loop checks that there’s room for one more byte, then writes two when it hits a ß. The output buffer gets 20 bytes of slack, so 21 or more ß’s overflow it. It’s reachable through <code>upcase</code>, <code>swapcase</code>, <code>downcase(:fold)</code> and <code>casecmp?</code>, and every Ruby back to <strong>2.4.0</strong> has it.</p>\n<p>Veria says a plain <code>rails new</code> app is reachable before login, but they haven’t said which input yet, and they plan to publish their exploit around November 25th. Ruby’s advisory says the team is “aware of PoCs that prove this flaw can be used to trigger RCE in some scenarios.” It credits Stephan Lipp of Ada Logics and Anthropic with the find.</p>\n<p>The fix is in the interpreter, so bumping Rails or any gem does nothing. Ruby 3.2 and older are end of life and won’t get a fix. And a Puma phased restart keeps the old Ruby running, so do a full restart. If you self-host GitLab or Mastodon, you need their updated packages.</p>\n<h2 id=\"php-round-two\" data-level=\"3\">PHP, Round Two</h2>\n<p>PHP shipped <strong>8.5.13</strong>, <strong>8.4.28</strong>, <strong>8.3.37</strong> and <strong>8.2.36</strong> on Thursday (along with 8.6.0), covering about 17 advisories, seven of them rated critical. It’s the follow-up to last month’s PHP2Shell, and the session upload progress feature shows up again.</p>\n<p><strong>CVE-2026-115560</strong> is the third RCE PHP reserved back in October. Upload progress data gets written into the session under one <code>session.serialize_handler</code> and read back with another, which hands attacker bytes to <code>unserialize</code>. It needs a non-default setup, like two apps sharing a session store with different handlers. The fix is the part to notice: <strong><code>session.upload_progress.enabled</code> is now off by default</strong>, so if something of yours shows upload progress bars from the session, they’ll stop working after the upgrade until you turn it back on.</p>\n<p>The other one to read about is <strong>CVE-2026-129289</strong>, request smuggling in PHP-FPM. Once a script opens <code>php://input</code>, PHP stops draining any body bytes the script doesn’t read. If the web server reuses FastCGI connections, those leftover bytes get parsed as the next FastCGI request, and a forged one can set <code>allow_url_include=1</code> and <code>auto_prepend_file=php://input</code> to run the body as PHP. Connection reuse isn’t the default, so check for it:</p>\n<figure class=\"code\" data-lang=\"bash\"><figcaption><span class=\"lang\">Shell</span></figcaption><pre class=\"shiki shiki-themes github-light github-dark\" style=\"--shiki-light:#24292e;--shiki-dark:#e1e4e8;--shiki-light-bg:#fff;--shiki-dark-bg:#24292e\" tabindex=\"0\"><code><span class=\"line\"><span style=\"--shiki-light:#6F42C1;--shiki-dark:#B392F0\">grep</span><span style=\"--shiki-light:#005CC5;--shiki-dark:#79B8FF\"> -rn</span><span style=\"--shiki-light:#032F62;--shiki-dark:#9ECBFF\"> \"fastcgi_keep_conn\"</span><span style=\"--shiki-light:#032F62;--shiki-dark:#9ECBFF\"> /etc/nginx/</span></span></code></pre></figure><p>If it says <code>on</code>, turn it off until you’ve upgraded.</p>\n<p>PHP2Shell itself is still causing damage. <a href=\"https://www.bleepingcomputer.com/news/security/check-point-gateways-hacked-via-php2shell-to-deploy-vmware-esxi-wiper/\">BleepingComputer reported</a> attackers getting into unpatched Check Point gateways through their PHP-based Mobile Access portal, using a second bug (CVE-2026-123039) for root, then moving to VMware ESXi hosts and running a wiper. If you run those gateways, look for session files containing <code>upload_progress_pwn</code>.</p>\n<h2 id=\"optimizely\" data-level=\"3\">Optimizely</h2>\n<p>Early Monday someone injected JavaScript into Optimizely Web Experimentation projects, and the A/B testing scripts those sites load from Optimizely’s CDN started serving a crypto address swapper and a card skimmer. <a href=\"https://www.securityweek.com/optimizely-vulnerability-exploited-to-serve-malware-on-major-websites/\">Report URI</a> checked 1,611 Optimizely scripts and found 188 infected, belonging to 178 organizations, including the BBC, CNN, Zoom, GitLab, Salesforce, HSBC and IKEA.</p>\n<p>The first injection they saw was at 04:21 UTC on the 16th, and everything was clean by 23:11 UTC. Optimizely says the attacker got in through a flaw in its experimentation account signup flow, then wrote a loader into each project’s custom JavaScript field. StepSecurity found that card numbers typed into third-party payment iframes were out of the skimmer’s reach, which is a good argument for those iframes.</p>\n<p>This is the problem with any script tag pointing at someone else’s server. Your page runs whatever is there today. A strict Content Security Policy won’t stop a script you’ve already allowed, but it can stop that script from sending data to an attacker’s domain with <code>connect-src</code>.</p>\n<h2 id=\"barracuda-and-artex-again\" data-level=\"3\">Barracuda And ARTEX Again</h2>\n<p>Remember ARTEX, the open-source agent pentest framework Microsoft tied to the ConnectWise attack last month? <a href=\"https://www.bleepingcomputer.com/news/security/barracuda-links-cloud-breach-to-artex-ai-powered-cyberattack/\">Barracuda says</a> it showed up in their cloud too. The way in was a cloud access key Barracuda had given an outside vendor for a past project and never retired. It was sitting in a config file in the vendor’s GitHub repo.</p>\n<p>Over 44 hours on November 1st and 2nd the attacker worked through about 10 Barracuda cloud accounts, reached source code, Kubernetes clusters and the system storing customer credentials for Barracuda’s XDR integrations, tried to launch a container that mounted the host’s file system, and tried to make itself a cloud admin (that part was blocked). Barracuda is asking affected XDR customers to recreate those credentials.</p>\n<p>Go look at your old access keys. The ones from finished projects are the dangerous ones, because nobody’s watching them.</p>\n<h2 id=\"qubole-and-expedia\" data-level=\"3\">Qubole And Expedia</h2>\n<p>Same shape, bigger blast radius. <a href=\"https://www.wiz.io/blog/qubole-cross-account-access-abused-in-the-wild\">Wiz found</a> that someone with AWS credentials belonging to Qubole, a big-data platform, spent October 14th through 26th assuming the integration roles Qubole customers had created in their own AWS accounts. Every request had the right External ID, so it all looked legitimate. Many of those roles had been sitting unused for a long time. Sentra counted about 100 secrets read from Secrets Manager, Parameter Store and Lambda environment variables in seven minutes.</p>\n<p>Expedia confirmed attackers reached it through “a compromised third-party data platform provider,” and a group calling itself n0n claims more than 830 million loyalty records. Expedia hasn’t confirmed that number.</p>\n<p>If you ever connected Qubole to AWS, search CloudTrail from October 1st for role sessions starting with <code>quiam-</code>, and do it soon. CloudTrail’s event history only keeps 90 days, so the October 14th activity starts disappearing in mid-January.</p>\n<h2 id=\"sckit\" data-level=\"3\">SCKit</h2>\n<p>This one started last Wednesday, but its second wave hit on the 13th. A third-party ScreenConnect extension called RMM+ (versions 1.0.31 and 1.0.32) accepted a blank shared key, and that key was blank by default. <a href=\"https://www.huntress.com/blog/sckit-worm-rogue-screenconnect\">Huntress watched</a> an attacker use it to skip ScreenConnect’s accounts and MFA entirely and push its own ScreenConnect client as SYSTEM to 288 machines at 48 companies in about 90 minutes, then 205 more machines at 41 companies two days later. ConnectWise disabled five extensions across roughly 400 partner environments. On-prem installs have to disable them by hand.</p>\n<h2 id=\"the-rest-of-the-week\" data-level=\"3\">The Rest Of The Week</h2>\n<ul>\n<li>CISA added a LiteLLM sandbox escape (CVE-2026-40217) to its Known Exploited Vulnerabilities list. The <code>/guardrails/test_custom_code</code> endpoint runs user Python in a homemade sandbox, and the default Docker image runs the proxy as root. It needs an admin key. Fixed in 1.83.11.</li>\n<li>uBlock Origin fixed a universal XSS where a malicious site could trick the element picker into saving a trusted filter that ran JavaScript on other sites. Update to 1.75.4 or newer.</li>\n<li>OpenSSL’s CVE-2026-121481 showed up with a scary CVSS 9.8 on the CVE record. OpenSSL itself rates it Low: it needs a Base64 input of about 1.6 GB, and nobody has found a way for an attacker to supply one. Fixes land December 1st in 4.0.4 and 4.1.1.</li>\n<li>Infoblox confirmed attackers got at company information through a third-party vendor.</li>\n<li>The Gentlemen ransomware group added NVIDIA to its leak site with no proof. NVIDIA says it’s looking into an incident at a third-party partner.</li>\n</ul>\n<p>Go restart your stuff, y’all. A reload won’t pick up most of this week’s fixes.</p>","created_at":"2026-11-21T15:00:00.000Z","updated_at":"2026-11-21T15:00:00.000Z"}