← Blog
November 20, 202620 min readThreat intelligenceAppSecCritical CVEsZero-daySupply chainRCE

Weekly AppSec Threat Digest: Optimizely Script Hijack, Axis RCE via Confluence, nginx HTTP/2 RCE

Nov 13–20, 2026: an attacker writes a crypto clipper into Optimizely-hosted scripts on 178 organizations’ sites, Apache Axis 1.x RCE is exploited through Confluence and added to KEV, Google ties a vCenter zero-day to ARMORY CASTLE, nginx and OpenJDK fix pre-auth memory corruption, and Microsoft’s agent-framework is backdoored on PyPI.

By The Xalgorix Team

On November 16, the JavaScript that large websites load from Optimizely to run A/B tests started swapping cryptocurrency wallet addresses. Nobody had poisoned a CDN or phished a password. An attacker used a flaw in Optimizely's account signup flow to write into customers' own experiment code, and Optimizely's CDN served the result to visitors of 178 organizations, by Report URI's count, for most of the day.

Much of the week looked like that: trust granted once and never checked again. Apache Axis 1.x builds a class named in a SOAP fault before any authentication runs, and attackers used it against Confluence Data Center days after disclosure; CISA added it to its Known Exploited Vulnerabilities (KEV) catalog on November 18. Google's threat intelligence team tied a vCenter authentication bypass to a September zero-day campaign. LiteLLM and REDCap joined KEV, and Apache says its Roller bug is being exploited. nginx and OpenJDK fixed pre-authentication memory corruption in HTTP/2 and in Base64 decoding, PHP, Drupal, WordPress and Spring shipped critical releases, and publishing credentials of still-unknown origin pushed a credential stealer into Microsoft's agent-framework package on PyPI.

At a glance (November 13–20, 2026): Optimizely Web Experimentation script injection (178 organizations, no CVE) · CVE-2026-90436 (Apache Axis 1.x, exploited via Confluence and N-central, KEV November 18) · CVE-2026-123890 (vCenter, exploited as a zero-day per Google) · CVE-2026-40217 (LiteLLM, KEV) · CVE-2026-90817 (REDCap, KEV) · CVE-2026-129448 (Apache Roller, exploited per Apache) · Puppet, Check Point, Synology and UnrealIRCd KEV additions · CVE-2026-123007 and CVE-2026-123633 (nginx HTTP/2) · CVE-2026-114938 (“Java4Shell”) · PHP, Drupal, WordPress 7.1.5 and Spring Security releases · the agent-framework and WPPConnect package compromises. As always, treat the linked vendor advisories as the operational source of truth.

Lead story: attackers write into Optimizely-hosted scripts on 178 organizations' sites

Optimizely Web Experimentation lets marketing teams change pages without a deploy. Each project has a Project JavaScript field that runs on every page where the Optimizely snippet loads, and the snippet is served from Optimizely's CDN. On November 16 an attacker wrote code into that field in many customers' projects. Optimizely confirms an attacker exploited a vulnerability in Experimentation and “used it to add malicious JavaScript to some customer projects,” and says the flaw was in the account signup flow and was used to impersonate users on targeted customer accounts. It has not explained how a signup flaw allowed that, and says it closed the hole within three hours of detecting it.

  • What ran. Scott Helme of Report URI counted 188 infected snippets across 178 organizations, including Zoom, the BBC, HSBC and GitLab, between about 04:21 and 23:11 UTC. The payload was rebuilt 26 times during the day. It replaced cryptocurrency wallet addresses in page text, form fields, the clipboard and outgoing requests, and sent cookies and a screenshot back after a swap. Report URI and StepSecurity also describe a card-skimming module added later in the day, which Socket's sample did not contain, and iPhone and iPad visitors were served loaders for the Coruna and DarkSword iOS exploit kits.
  • Why most CSPs did not help. Report URI checked the affected sites' Content Security Policies. Allowlisting Optimizely's CDN did nothing, because the first stage came from it. Only 24 organizations blocked the payload, 68 had no CSP at all, and 17 let it run but stopped the stolen data from leaving with connect-src. Policies with 'strict-dynamic' or 'unsafe-eval' let it through: the first extends trust to whatever a trusted script loads, and the payload carried an eval fallback for the second.
  • Where it stands. Optimizely's November 18 update says no malicious code remains in its systems and that fewer than 5% of customers were affected.
Signup-flow flaw
Used to impersonate users
Project JavaScript
Edited in customer projects
Vendor CDN
Allowlisted host serves it
Visitors
Wallet swap, cookies, skimmer
A trusted host serving attacker code. Based on Optimizely's status page and Report URI's analysis.

Remediation: if you use Optimizely Web Experimentation, review the Project JavaScript and other custom-code fields in every project, review users and API tokens, and search logs back to 04:00 UTC on November 16 for the published indicators, including requests to 1688360[.]cc (track.js, hit.php, card.php) and the window._k7 page marker. One loader variant base64-encodes the domain, so a plain string search can miss it. Treat card and wallet details entered on affected pages that day as exposed.

The longer lesson is that anyone with write access to a tag manager or experimentation tool can ship code to your production pages. Keep those tools off checkout and payment pages (PCI DSS 4.0 requirements 6.4.3 and 11.6.1 already ask for an inventory of payment-page scripts and change detection on them), monitor what third-party scripts contain rather than only where they come from, drop 'unsafe-eval', and set a tight connect-src, the directive that stopped exfiltration on 17 sites. Our guide to deploying a strict Content Security Policy recommends 'strict-dynamic' for first-party scripts; this incident shows its limit, because a compromised vendor snippet inherits that trust. See Optimizely's status page, Report URI's analysis, Socket and StepSecurity.

Actively exploited this week

CVE-2026-90436 — Apache Axis 1.x is exploited through Confluence Data Center (KEV)

Last week this bug reached N-central; this week it reached Confluence. When an incoming SOAP fault names an exception class, Apache Axis 1.x loads and instantiates that class before checking that it is an AxisFault, and before any per-service authentication runs. Loading a suitable class runs its static initializer, which is enough for code execution. Confluence Data Center embeds Axis for its Remote API. Rapid7's MDR team saw exploitation on November 13: unauthenticated POSTs to /plugins/servlet/soap-axis1/confluenceservice-v2, the Confluence process launching bash and curl within seconds, the attacker using Confluence's database credentials about three minutes later, a service-account credential and queries against a cloud data warehouse within about four hours, and a Microsoft 365 sign-in with an integration account after about twelve. A Metasploit module followed on November 17, and CISA added the CVE (CVSS 9.8) to KEV on November 18 with a November 21 deadline.

Remediation: Confluence Data Center 9.2.28, or 10.x, which removed the feature. Otherwise disable the Remote API (XML-RPC & SOAP) and confirm the servlet now answers 403; Confluence Cloud is not affected. For N-central, 2026.4 HF3 (build 2026.4.1.39), released November 16, removes the Legacy API and fixes 17 further CVEs (CVE-2026-129271 through CVE-2026-129287), and N-able says every agent and probe must be upgraded too. GreyNoise has logged 13 successful intrusions against its N-central decoys since November 9: the first on November 11, using default credentials and a file-upload bug on a decoy older than HF2; the first through the Axis bug on November 15, against a decoy running HF2; and none against HF3. Axis 1.x itself is end of life with no fixed release planned, so any other product that bundles axis-1.4.jar needs a migration or a patched build. See Atlassian's advisory, Rapid7, the Apache advisory, N-central HF3 release notes and GreyNoise's tracker.

CVE-2026-123890 — a vCenter authentication bypass that Google says was a September zero-day

Broadcom's VMSA-2026-0009 (November 19, CVSS 9.8) fixes an unauthenticated flaw in the VMware Directory Service: over DCE/RPC, an attacker can obtain an SRP-6a session key and act as administrator@vsphere.local. On November 20, Google Threat Intelligence Group reported that a suspected PRC-nexus actor it tracks as ARMORY CASTLE (formerly UNC5221) exploited it in early September with a tool it calls BONESAW, after getting in through a BeyondTrust Remote Support zero-day (CVE-2026-102764). GTIG describes webshells, a new SSO administrator account, theft of the directory database and SAML signing keys, which is enough to forge sign-ins (“Golden SAML”), and access to ESXi hosts.BleepingComputer notes that Broadcom had not itself confirmed exploitation, and the CVE was not in the November 18 KEV catalog.

Remediation: vCenter 9.1.1.0200, 9.0.2.0200 or 8.0 U3m (VCF 5.x through KB 88287); Broadcom lists no workaround. vCenter 7.0 is also affected and gets a fix only through extended support, and GTIG says BONESAW works against both 7.x and 8.x. Hunt for unexpected SSO administrator accounts and for SSH being enabled through the appliance API. If you find a compromise, rotate the SAML signing certificates, because stolen signing keys keep working after the patch. See Broadcom's advisory and GTIG's report.

CVE-2026-40217 — LiteLLM's guardrail tester turns into an admin key (KEV)

LiteLLM, the open-source gateway many teams put in front of their LLM providers, exposes POST /guardrails/test_custom_code for trying out custom guardrails. It runs user-supplied Python in a homemade sandbox that filters the source text; building blocked names at runtime and rewriting bytecode gets back to unrestricted builtins, and the default Docker image runs the proxy as root. X41 D-Sec disclosed it in April. On November 17, GTIG reported a threat actor exploiting it, describing a sandbox escape that lets an attacker “elevate limited-access inference keys (‘Virtual Keys’) into full administrative credentials,” and CISA added it to KEV on November 18. The project's advisory says that in default configurations the endpoint requires a proxy-admin credential. Scores range from 5.9 (the project's own CVSS v4 rating) to 8.8 (the CVE record's CVSS v3.1).

Remediation: GitHub's reviewed advisory lists 1.83.10 as the first fixed release while the project's text says 1.83.11, so go to 1.83.11 or later. Block the endpoint at your proxy, replace default or guessable master keys (Zenity's honeypot saw attempts with keys such as sk-1234), and run the container as a non-root user. A gateway that holds every provider key deserves the care you give an identity provider; the controls in our piece on least privilege for AI agent tools apply here too. See the LiteLLM advisory, X41's write-up and GTIG.

CVE-2026-90817 — REDCap survey links reach the data importer (KEV)

REDCap, the research-data platform from Vanderbilt University, let a request made in a public survey's context be routed to its data import controller through a passthrough parameter, where a crafted import ran code. An attacker needs only a valid public survey hash. Unit 42 reported cryptominer deployments in October, with a Blade-template webshell under temp/cache followed by XMRig, and CISA added the CVE (CVSS 9.8) to KEV on November 18 with a November 21 deadline. Remediation: upgrade to 16.0.49 or 17.3.10 (LTS) or 17.4.4 (Standard), then search access logs for __passthru requests that reach DataImportController. See the NVD entry and Unit 42's note.

CVE-2026-129448 — Apache Roller says its XML-RPC bug is being exploited

Apache Roller's November 17 advisory fixes a pre-authentication deserialization flaw (CVSS 9.8): the XML-RPC endpoint deserializes vendor extension types before it authenticates the caller. Apache says “This issue is being actively exploited,” without details. Versions 4.0.0 through 6.1.4 are affected, 6.1.5 is covered by the same flaw under the earlier CVE-2026-82384, and Apache recommends 6.1.6 or later. If you cannot upgrade, set enabledForExtensions=false in WEB-INF/web.xml or remove the /roller-services/xmlrpc mapping and restart; turning XML-RPC off in the admin UI is not enough. It is the same order of operations as Axis: parse untrusted types first, authenticate second. See the Apache advisory.

Also added to KEV this week

  • Puppet Enterprise and Puppet Server CVE-2026-115731 (CVSS v4 9.4) — a report sent to PUT /puppet/v3/report can carry a __ptype that reaches class_eval, running code as the Puppet Server user. Any CA-signed agent certificate is enough, and with unverified autosigning anyone can get one. Fixed in PE 2023.8.13, 2025.12.1 and 2026.0.1 and in OpenVox Server 8.18.0 and 9.2.0; review autosign settings and the certificates you have signed. Added November 17. (OpenVox advisory)
  • Check Point Gaia CVE-2026-123039 (CVSS 7.8) — the root dynamic_server daemon accepted unauthenticated commands on a world-writable socket at /tmp/dynamic_socket. Check Point's advisory says it observed exploitation against “a couple of dozen customers,” and its research team describes attackers getting in through the PHP2Shell bugs in the Mobile Access portal, then deploying backdoors, webshells and in some cases a wiper. Install the fixed Jumbo Hotfix takes (R82.20 Take 14, R82.10 Take 52, R82 Take 132, R81.20 Take 174 or later); R81.10 and R81 are end of support and get no fix. Added November 13. (sk1000362, Check Point Research)
  • Synology DSM CVE-2024-10441 (CVSS 9.8) — a 2024 Pwn2Own bug in DSM's system plugin daemon. Synology warned on October 29 that ransomware is hitting outdated DSM versions; which CVE that campaign uses is not confirmed. Update DSM and turn on automatic updates. Added November 18. (Synology-SA-24:20)
  • UnrealIRCd CVE-2026-114713 — the IRCv3 label tag was echoed into replies without escaping, so a user on a network of linked servers could inject server-to-server commands. Upgrade to 6.2.7.1 or apply the hot-patch. Added November 13. (oss-security)

Critical web application CVEs (no confirmed exploitation)

CVE-2026-123007 and CVE-2026-123633 — nginx HTTP/2 remote code execution

nginx 1.30.7 and 1.31.9, released November 18, fix two HTTP/2 memory-safety bugs that Wiz Research and Xint each say lead to code execution in the worker process without authentication. In CVE-2026-123007, from code dating to 2015, nginx refunds the send window for cancelled queued data without capping it at the HTTP/2 maximum; an unsigned bounds check then wraps, and the result becomes a negative length for a copy in the TLS send path. Wiz says it demonstrated code execution over HTTPS on 64-bit builds against a server serving only static files, with no backend or third-party module. CVE-2026-123633 is a use-after-free: a connection still waiting for the rest of a HEADERS frame stays on the reusable list, so when worker connections run short nginx can free it while a proxied request is still using it. F5 scores both 9.2 under CVSS v4 (8.1 under v3.1); nginx.org rates the first “major” and the second “medium.”

Cancel queued streams
Window refunds add up
Uncapped send window
No HTTP/2 maximum check
Unsigned check wraps
Length goes negative
Out-of-bounds copy
TLS send path, worker RCE
Simplified from the Wiz and Rapid7 write-ups.

Rapid7's Stephen Fewer has published a working exploit for CVE-2026-123007, with real limits: it needs HTTP/2 over TLS, a glibc whose memcpy takes the AVX-512 path, and about 128 GB of HTTP/2 traffic, and it was shown against a freshly started container. No one reports in-the-wild exploitation. Wiz credits five teams with finding the bugs independently. The same releases also fix use-after-frees in the mail proxy (CVE-2026-113425), HTTP/3 (CVE-2026-121810) and the resolver (CVE-2026-123634), each of which Xint says reaches code execution on its own, a TLS 1.3 request-body loop (CVE-2026-118899) and several more. Remediation: upgrade nginx, or take the NGINX Plus releases in F5's advisories; distributions may backport the fix without changing what nginx -v prints. As a stopgap, turn HTTP/2 off (http2 off; or remove http2 from listen), which does not cover the mail, HTTP/3 or resolver bugs. Watch for long-lived HTTP/2 connections with unusual numbers of WINDOW_UPDATE frames, and for worker crashes. See nginx security advisories, Wiz, Rapid7 and Xint.

CVE-2026-114938 — “Java4Shell,” a Base64 decoder that writes past its buffer

On x86-64 CPUs with AVX-512 VBMI, the JIT-compiled Base64 decoder in OpenJDK 17.0.6 and later computes the wrong store mask for a short padded tail: the length underflows and the decoder makes a 64-byte write that runs up to 62 bytes past the end of its output array. Once the decoder is hot, any Base64 a server decodes before authentication can reach it. Calif, which published the research, says it has verified pre-authentication code execution in default configurations of Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate; it is withholding those chains, and its public trigger corrupts memory without running code. Oracle scores it 8.1, and CISA issued an alert on November 17 urging immediate patching. GreyNoise had logged 197 Java4Shell-shaped requests from seven IPs by November 20, with no confirmed successful exploitation. ARM64 is not affected.

Remediation: OpenJDK 17.0.21.1, 21.0.13.1 or 25.0.5.1, or your vendor's equivalent build, including JREs bundled inside products and container images. Until then, start the JVM with -XX:-UseBASE64Intrinsics and, as Keycloak advises, confirm the flag took effect with jcmd <pid> VM.flags -all. See Calif's write-up, CISA's alert and Keycloak's guidance.

PHP 8.6.0 and 8.5.13 — FastCGI request smuggling and a session race

PHP's November 19 releases (8.6.0, 8.5.13, 8.4.28, 8.3.37 and 8.2.36) fix several issues the php-src advisories rate Critical (CVSS v4 9.2), although PHP notes they “require either non-default configuration or uncommon setups.” Two matter most for web teams. CVE-2026-129289 is FastCGI request smuggling: when an application opens php://input without reading it to the end, PHP skips discarding the unread body and PHP-FPM loses its place inside the FCGI_STDIN record, so a forged FastCGI request hidden in the body gets processed, running code as the FPM worker. It needs a front end that reuses backend connections (nginx fastcgi_keep_conn on, Apache enablereuse=on, some HAProxy setups). CVE-2026-126714, which its finder, depthfirst, named “phpAfterShell,” dates back to PHP 4.1.0: the default files session handler overwrites a session file without truncating it, so an interrupted write leaves a mix of old and new data that the next session_start() deserializes. Upload progress is the main way in, and the application still needs a usable gadget chain. No exploitation has been reported.

Remediation: upgrade, rebuild images and restart FPM. Until then, turn off FastCGI connection reuse and, if you do not use it, set session.upload_progress.enabled = 0. See the php-src advisories, oss-security and Wiz.

Drupal SA-CORE-2026-017 — the installer as a way back in

Drupal's November 18 core release fixes CVE-2026-132165, rated Critical (18 of 25): under conditions Drupal does not spell out, an anonymous visitor can use install.php to reinstall the site, leading to takeover or code execution. Drupal's risk rating says a proof of concept exists, though no exploitation is reported, and some paths fail when the web server cannot write to settings.php. Fixed in 10.6.22, 11.3.22 and 11.4.13; 11.2.x, 10.5.x and older are end of life and get no fix. A companion advisory, SA-CORE-2026-018, removes a deserialization gadget chain that matters only if a site has a separate insecure-deserialization bug. Remediation: upgrade, lock down the installer as Drupal's documentation describes, and make sure settings.php is not writable by the web server. See SA-CORE-2026-017.

WordPress 7.1.5 — 20 fixes, one unauthenticated path toward code execution

WordPress 7.1.5 (November 18) is a security release; WordPress says backports to every branch back to 4.7 are in progress. The headline fix is an “unauthenticated privilege escalation via Template Part block attributes that could lead to remote code execution,” reported by Logan Homolka; Patchstack reads the diff as a REST path injection whose route to code execution runs through the plugin-install endpoints when a privileged user opens the content. Other fixes include SQL injection in WP_Meta_Query through nested relation values (only where a plugin or theme passes request data into meta_query), unauthenticated stored XSS through comment shortcode attributes, contributor-level stored XSS in the Comment Date block, a multisite Administrator-to-Super-Admin escalation and an XML-RPC pingback denial of service. No CVEs had been assigned when Patchstack wrote its analysis, and nothing is reported exploited. Remediation: confirm that every site actually received the background update. See the release announcement and Patchstack.

Spring Security CVE-2026-59273 and Spring AI filter injection

Spring's November 19 batch of advisories includes a Critical one in Spring Security: ELRequestMatcher evaluates SpEL in an unrestricted context, and because Spring's data binding finds a property editor for the type automatically, external input bound to a matcher property can become an expression with reflection and arbitrary method calls. Spring does not describe the preconditions further, but in practice an application has to bind request input to an ELRequestMatcher-typed property. Fixed in open-source 7.1.3 and 7.0.9; 6.5.13, 6.4.20, 5.8.29 and 5.7.27 are available only through enterprise support. The same day, Spring AI 2.0.3 fixed vector-store filter bugs, including RediSearch query injection in RedisVectorStore (CVE-2026-59361) that can break tenant isolation. If user- or model-supplied values feed vector-store metadata filters, those filters are not an authorization boundary. Neither advisory mentions exploitation. See CVE-2026-59273 and CVE-2026-59361.

Control Web Panel — a recovery code that doubles as an API key

VulnCheck's Valentin Lobstein published two chains on November 16 against Control Web Panel, a free Linux hosting panel with about 124,000 internet-facing instances by VulnCheck's count. CVE-2026-124020 is classic Host-header poisoning: the public lostpass endpoint builds the emailed reset link from the request's Host header, so once the account owner or a link-following mail scanner opens it, the attacker receives the recovery code, logs in, and gets a shell as the hosting account through the File Manager or Crontab. In CVE-2026-124021, the admin API's validate() accepted a password-recovery code as authorization, and the legacy signature check relies on a hostname that the TLS certificate on port 2304 gives away; the result is command execution through cron jobs. VulnCheck wrote working exploits and does not report in-the-wild use. Fixed in CWP 1.30.6. Build security-relevant URLs from a configured canonical host, never from request headers, and make recovery codes single-use and single-purpose. See VulnCheck's write-up.

Also patch this week

  • GitLab 19.5.6, 19.4.8 and 19.3.9 (November 16) — 37 CVEs, four High: stored XSS in wiki pages and in Error Tracking (both 8.7), an authorization race that exposes protected CI/CD variables (CVE-2026-18241, 8.2) and a Workspaces flaw that reads Kubernetes Secrets from other namespaces (CVE-2026-97187, 7.6). Also notable: OAuth tokens scoped to MCP could approve and merge merge requests (CVE-2026-123142). (GitLab)
  • Gitea 28.4.0 — a push during repository auto-initialization plants a README.md symlink that overwrites the post-receive hook, running the repository description as a script (CVE-2026-125423, 8.5, not stopped by DISABLE_GIT_HOOKS), and /repo/search returned the password hashes of public-repository owners to any signed-in user (CVE-2026-123006, 8.8, so force password resets). (GHSA-6m4f-m9vc-7946, GHSA-frjx-8pcw-8c9q)
  • SmarterMail Build 9735 — application passwords generated with too little entropy allow a pre-authentication bypass (CVE-2026-126969, 9.1), and an authenticated File Edit path traversal gives code execution as SYSTEM (CVE-2026-126967, 8.8). (ZDI-26-997, ZDI-26-994)
  • Ruby 4.0.10, 3.4.13 and 3.3.14 — seven memory-safety bugs in Regexp compilation and String case conversion. Ruby says proofs of concept reach code execution “in some scenarios” for five of them and show attacker-controlled memory corruption for a sixth; it does not say whether they are public. Audit code that compiles user-controlled patterns. (ruby-lang.org)
  • Apache Shiro 3.0.1 and Apache APISIX 3.20.0 — with an LDAP realm and authentication caching (off by default), Shiro accepted any password after a user's first login (CVE-2026-129349); APISIX route-level authentication could be bypassed with encoded dot segments (CVE-2026-124680). (Shiro, APISIX)
  • Eclipse Mosquitto 2.1.4 — pre-auth code execution in the built-in WebSockets listener (CVE-2026-110486), relevant to MQTT-over-WebSocket dashboards. (oss-security)
  • AI tooling CVEs that reached NVD this week — Arize Phoenix 15.7.0 fixes a format-string injection that leaks the signing secret on default installs, which have authentication off (CVE-2026-53998, CVSS 10.0), and the MCP Python SDK 1.30.0 and 2.2.0 stop a malicious MCP server from steering the OAuth client's token request to an endpoint it controls (CVE-2026-115563). Both GitHub advisories were published earlier; their CVE records reached NVD this week. (GHSA-gfv3-m769-5xm3, GHSA-qx49-fqc8-xw99)

Follow-ups

  • Apache httpd CVE-2026-119618. Singapore's Cyber Security Agency says the mod_http2 bug from our November 6–13 digest is “reportedly being actively exploited.” Nobody else corroborates that, it is not in KEV, and Sysdig says the public proof of concept corrupts the heap without running code. Upgrade to 2.4.71 either way. (CSA AL-2026-164)
  • SSHole, Barracuda and wfx. We found no material change to OpenSSH's SSHole (CVE-2026-123158): no new OpenSSH release, and GreyNoise's tracker shows no new attack chain this week. NVD now scores Barracuda's CVE-2026-116606 at 9.8 and CVE-2026-111560 at 7.2. The Aikido and StepSecurity write-ups on the wfx GitHub Actions secret stealer have not been updated since November 12.

Supply chain: the publisher looked legitimate

Microsoft agent-framework 1.26.1 and 1.26.2 on PyPI

On November 17, two malicious releases of agent-framework, Microsoft's Python SDK for building AI agents, appeared on PyPI about half an hour apart (20:35 and about 21:08 UTC) and were gone by about 21:30. They shipped a .pth file (agent_framework_meta.pth, then af_runtime_diag.pth), which Python executes at every interpreter start, so the stealer ran wherever the package was installed, imported or not. It collected environment variables, including other processes' through /proc/*/environ, and credential files for Azure, Azure DevOps, npm, git and Docker; requested fresh Azure managed-identity tokens; and, in 1.26.2, requested GitHub Actions OIDC tokens in workflows with id-token: write, sending everything to a webhook.site URL. Wiz counted three successful exfiltrations and assesses that the attacker “appears to have compromised a client certificate and private key” for Microsoft's ESRP release-signing pipeline; Aikido says only that valid publishing credentials were used. We found no Microsoft statement.

Remediation: pin 1.26.0 and rebuild any environment, CI cache or image that pulled either version; look for unexpected .pth files in site-packages; rotate everything those machines could reach; review sign-ins and federated credentials for the pipeline identities; and search egress logs for webhook.site. As our article on trusted publishing and release hardening argues, a correctly published package tells you which pipeline shipped it, not whether that pipeline was under its owner's control. See Wiz, Aikido and Socket.

WPPConnect: unsigned commits, released by the project's own pipeline

StepSecurity reports that on November 15 and 16, unsigned commits were pushed straight to master on the WPPConnect WhatsApp projects through a maintainer's account, and the project's release automation published them with valid trusted-publishing credentials: npm @wppconnect-team/wppconnect 2.4.1 and 2.4.3, @wppconnect/server 2.10.82 and 2.10.83, and the matching Docker images, including the latest, 2 and 2.10 tags during the window. The branches were later force-pushed to hide the commits. The payload runs on import or server start and sends environment variables, .env and config files, and WhatsApp session tokens to a Telegram bot. This rests on a single source, and as of November 17 the maintainers had not confirmed that the publishing account was contained or the Telegram credentials revoked. Remediation: pin 2.4.0 and 2.10.81, pull Docker images by digest, rotate secrets and re-pair WhatsApp devices. Branch protection that requires reviewed, signed commits would have stopped the first step. See StepSecurity.

Web application and AI security research worth your time

XSS in innerHTML without an equals sign — Masato Kinugawa, PortSwigger Research, November 18

A script element inserted through innerHTML never runs; that is the rule sanitizers lean on. Kinugawa shows two browser behaviours that combine to break it. In Chromium, a script element created inside SVG this way was left without its “already started” flag, and the new <selectedcontent> element clones the children of the selected <option> into itself. Put <svg><script> inside an option of a <select> that contains <selectedcontent>, and the clone runs, with no = anywhere in the payload, so filters that look for attributes or event handlers see nothing. The idea grew out of a May 2026 DOMPurify bypass by Kabir Acharya (GHSA-87xg-pxx2-7hvx). Chrome changed the behaviour in version 153; the article says Safari 27.2 is still affected. Practical takeaway: character and keyword blocklists are not an XSS defence. Encode for the output context, use Trusted Types where you can, and re-test sanitizer configurations whenever browsers ship new elements. Our guide to telling exploitable XSS from a harmless popup covers what to do once a payload fires. Read the research.

Escaping containers through Docker Model Runner — ZDI, November 18

Nitesh Surana shows that any container could reach model-runner.docker.internal, and that runtime flags sent to its /engines/_configure endpoint were appended to llama-server's command line. Flags such as --log-file, --grammar-file and --chat-template-file then give file read and write on the host (ZDI-26-150, CVE-2026-28400), and with a Docker Extension installed that becomes code execution on the host. The post also covers a macOS sandbox escape and a WSL2 pivot that Docker treats as outside its threat model. Fixed in Docker Model Runner 1.0.16 and Docker Desktop 4.61.0. Practical takeaway: local model-serving APIs are attack surface; authenticate them, and allowlist flags instead of passing them through. Read the research.

When the guard model is weaker than the model it guards — Check Point, November 19

Check Point Research describes “PuzzleMask,” which hides instructions in ordinary prose, with no invisible characters or formatting tricks, so that a lightweight guard classifier passes the text while the stronger downstream model, with reasoning and a code interpreter, decodes it and follows it. The write-up reports success against one frontier model and a block by another, without success rates. The structural point is the useful one: if a cheaper model screens input for a more capable agent, the agent will understand things the screen cannot. Practical takeaway: size the agent's tool permissions and data access so that it stays safe when the guard misses. Read the research.

Disclosure clocks are getting shorter

Google Project Zero's 2026 policy update (November 13) replaces “90+30” with a flat 90 days now, 60 days from February 2, 2027 and 30 days from January 3, 2028, with no grace periods, citing AI speeding up both discovery and exploitation. On November 19, Google's Vulnerability Reward Program said reports submitted after December 1, 2026 that earn $10,000 or more will be published automatically at least 30 days after the fix. Chrome's security team also described “Mirror Dimension,” a sealed synthetic web for testing browser agents against prompt injection through rendered page visuals, DOM elements and WebMCP tool interactions. For defenders the message is the same: the gap between a fix and a detailed public write-up is shrinking. See Project Zero, the VRP update and Mirror Dimension.

Bug bounty highlights

Basecamp: a YouTube embed that makes an outside collaborator an admin — HackerOne #4175245

Disclosed on November 13, matty69v's High-severity report chains several steps. An outside collaborator posts a message that renders as a normal YouTube embed, with HTML hidden, encoded, in the bc-attachment content attribute. When an admin opens the message's edit form, the Lexxy editor re-normalizes select > option > div structures, and data-behavior and data-url attributes survive DOMPurify. A legacy load_drafts_link behaviour then fetches attacker HTML into innerHTML, and an injected cross-origin turbo-frame gets the page's CSP nonce copied onto its script. That script reads the CSRF token and grants the attacker admin through POST /account/administratorships. Testing takeaway: sanitize after the last transformation, not before it; look for behaviours driven by data-* attributes in old JavaScript; and check whether your framework copies CSP nonces onto remotely loaded content. See the report.

Weblate: a backup import that skips the role checks — HackerOne #4081135

In zukosdead's Critical report, disclosed November 15, ProjectBackup.restore_team() restored team roles without the checks the UI and API apply, and global permissions did not filter on the defining project, so a low-privilege user allowed to import backups gained site-wide permissions from an edited weblate-backup.json. Testing takeaway: every import, restore and migration path needs the same authorization as the front door; add them to the endpoints you test in our guide to IDOR and BOLA. See the report.

Two more worth a look

  • Essity, HackerOne #4191812 (disclosed November 18) — an unauthenticated HTML-to-PDF endpoint rendered attacker HTML with ABCpdf's old Gecko engine, so an iframe pointing at file:///C:/Windows/win.ini came back inside the PDF. Server-side renderers need network and file: access turned off. (report)
  • PortSwigger, HackerOne #4000535 ($1,000, disclosed November 13) — Burp's OpenAPI action validated one object in a response and acted on another, and a file: URL in $id rebased a fragment-only $ref past the external-definition checks. Parse once, then pass the parsed object along. (report)

The pattern: trust checked once, somewhere else

Optimizely's CDN was trusted, so its content was too, and 'strict-dynamic' passes a script's trust to whatever it loads. Axis instantiates a class before checking its type or the caller, and Roller deserializes before it authenticates. DOMPurify cleaned Basecamp's HTML before an editor rewrote it. Weblate's importer and Burp's second parser skipped checks the main path enforces. Control Web Panel trusted the Host header and took a recovery code as an API credential. agent-framework and WPPConnect were published with credentials and pipelines that were real but no longer under their owners' control. And PuzzleMask slips past a guard that understands less than the model behind it.

In every case the decision was made on an earlier or weaker view of the input than the one that finally acted on it. Reading either component alone rarely shows the gap, and matching versions or response patterns does not either. What shows it is sending the input the last component will actually see and checking what it does, which is the case for testing by exploitation rather than by inference.

What to do first: November 13–20 action matrix

PriorityTarget / CVESeverity & impactAction
P0 — todayOptimizely Web Experimentation customersInjected crypto clipper and skimmer · Exploited November 16Review Project JavaScript, users and tokens; search logs for 1688360[.]cc; tighten connect-src.
P0 — todayConfluence Data Center, N-central, Axis 1.x — CVE-2026-90436CVSS 9.8 · Pre-auth RCE · KEV, exploitedConfluence 9.2.28 or 10.x, or disable the Remote API; N-central HF3 plus agents; inventory axis-1.4.jar.
P0 — todayvCenter — CVE-2026-123890CVSS 9.8 · SSO admin without credentials · Exploited per GTIG9.1.1.0200, 9.0.2.0200 or 8.0 U3m; hunt for new SSO admins; rotate SAML signing keys if compromised.
P0 — todayLiteLLM — CVE-2026-40217 · REDCap — CVE-2026-90817 · Apache Roller — CVE-2026-129448Admin takeover / pre-auth RCE · ExploitedLiteLLM 1.83.11+ and block the guardrail tester; REDCap 16.0.49 / 17.3.10 / 17.4.4; Roller 6.1.6.
P0 — todayagent-framework 1.26.1 / 1.26.2 · WPPConnectCredential stealers in published packagesPin 1.26.0, 2.4.0 and 2.10.81; check for stray .pth files; rotate secrets and CI tokens.
P1 — this weekPuppet · Check Point Gaia · Synology DSM · UnrealIRCdKEV additions · RCE / local rootPE 2023.8.13 / 2025.12.1 / 2026.0.1; fixed Jumbo takes; current DSM; UnrealIRCd 6.2.7.1.
P1 — this weeknginx — CVE-2026-123007 / 123633 · OpenJDK — CVE-2026-114938Pre-auth memory corruption · Public PoC (nginx) · Probing (Java)nginx 1.30.7 / 1.31.9 or HTTP/2 off; patched JDKs or -XX:-UseBASE64Intrinsics.
P1 — this weekPHP · Drupal · WordPress · Spring Security · Control Web PanelCritical releases · Not exploitedPHP 8.6.0 / 8.5.13 / 8.4.28 / 8.3.37 / 8.2.36; Drupal 11.4.13 / 11.3.22 / 10.6.22; WordPress 7.1.5; Spring Security 7.1.3 / 7.0.9; CWP 1.30.6.
P1 — this weekGitLab · Gitea · SmarterMail · Ruby · Shiro · APISIX · MosquittoHigh and Critical fixes · Not exploitedGitLab 19.5.6; Gitea 28.4.0 and reset owner passwords; SmarterMail 9735; Ruby 4.0.10; Shiro 3.0.1; APISIX 3.20.0; Mosquitto 2.1.4.
P2 — studyThird-party script CSP · <selectedcontent> XSS · import paths · guard modelsResearch — testing coverageMonitor vendor script content; re-test sanitizers; give import and restore the same authorization; bound agents without the guard.

Sources and further reading

Ready to see it prove a bug?

Start scanning on Cloud →