Five *2Shells in Five Weeks: Read the Prerequisites Before the Headline

PHP2Shell, Node2Shell, Java4Shell, NGINX2Root and Ruby2Shell are all real, but not equally urgent. What each one needs, and how to rank your patching by exposure.

Conviso Research Team

November 22, 2026

8 min de leitura

TL;DR

  • Between 21 October and 19 November, five widely used runtimes and servers got a “*2Shell” or “*2Root” bug: PHP2Shell, Node2Shell, Java4Shell (Java2Shell), NGINX2Root and Ruby2Shell. All five are real memory-safety bugs, found by serious researchers, and all five deserve a patch.
  • They are not equally urgent for everyone. Each one needs a specific set of conditions: a CPU feature, a legacy text encoding, ~128 GB of HTTP/2 traffic, a particular SAPI, or a body parser that decompresses Brotli. Whether those conditions hold is a fact about your environment, not about the CVE.
  • Ordering the work by evidence and exposure rather than by name or CVSS score is what a mature assessment does, and it’s where specialist support pays off.

Real Bugs, Different Risks

This post isn’t arguing that these bugs are overblown. Each one is a genuine flaw in software that runs a large share of the internet, and in several cases researchers showed working remote code execution against real products.

The point is narrower. A headline that says “critical pre-auth RCE in Java” is a statement about the worst case. Every one of these advisories also lists the conditions that make the worst case possible, and those conditions are usually in the third paragraph, the researcher’s blog, or the PoC README. When you put them side by side, the five bugs look very different from each other, and the question “how worried should we be?” gets an answer you can act on.

So each bug below gets the same card:

  • Needs: what has to be true for exploitation, as documented by the vendor or the researchers.
  • Realistic worst case: what most affected deployments should actually expect.
  • Exploitation status: as of 19 November 2026.
  • Name check: where the nickname and the facts don’t quite line up.
  • Check first: the quickest way to find out whether the card applies to you, and what the stopgap does not cover.

The cards are in disclosure order.

1. PHP2Shell (CVE-2026-109618 / CVE-2026-109619)

Disclosed 21 October. Fixed in PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35. (How the two bugs work and how PHP redesigned its out-of-memory handling to fix them is covered in our earlier PHP2Shell analysis. This card is only about exposure.)

Needs

  • mod_php has the most direct path. Remote code execution “was validated over HTTP against mod_php with default PHP input and memory limits and normal prefork configurations with multiple workers” [1].
  • PHP-FPM is exploitable too, but it depends on the application.
    • PHP’s advisory for CVE-2026-109619 says “Session lifecycle RCE was validated against mod_php and FPM”. It adds that the FPM path “works with a normal dynamic FPM pool, although it depends on a suitable application operation” [2].
    • Calif, one of the original reporters, describes “reliable exploits against mod_php and PHP-FPM” and has a demo running against nginx with PHP-FPM. The operation it needs after startup is one “such as session_start(), session_decode() or unserialize()” that the attacker can reach. Calif calls it “application-dependent RCE” [3]. Session handling is common in any PHP application with logins.
    • Trail of Bits reports shells on MediaWiki, Symfony, Laravel and phpBB, and says its exploits work on pre-patch PHP 8.2–8.5 [39].
    • The line that’s often quoted, that “no complete FPM exploit has been demonstrated”, comes from the CVE-2026-109618 advisory and covers only that out-of-memory path [1].
  • Worker reuse affects reliability, but doesn’t stop the attack. “Multiple workers and recycling reduce success per request but do not remove the condition” [2].
  • A default that’s on. “Session upload progress is enabled by default, and no application call to session_start() is required” [2].

Realistic worst case

  • Unauthenticated RCE on mod_php.
  • RCE on PHP-FPM where the application performs a session or unserialize() operation the attacker can reach.
  • Unauthenticated RCE on appliances whose PHP-backed admin or VPN portal is reachable from the internet. That’s where confirmed compromises happened.

Exploitation status: exploited. CISA added both CVEs to the KEV catalogue on 26 October [5].

  • Among the confirmed victims were network appliances with exposed management or portal interfaces, some CTF-style websites, and others [6][40].
  • Palo Alto puts the share of its installed base with internet-facing management interfaces at “approximately .02%” [7].
  • Check Point reports a couple of dozen compromised gateways through Mobile Access portals between 30 October and 4 November [8].

Name check: one name covers two CVEs. They were patched together, but they have different exploitation paths and different SAPI exposure (see Needs).

Check first

  • Check Server API and session.upload_progress.enabled as the web server sees them: use php-fpm -i | grep -E '^Server API|session.upload_progress.enabled'. php -i on the command line reports the CLI SAPI and the CLI’s own php.ini.
  • Then the step that’s easy to miss: list every appliance and product that bundles its own PHP (firewalls, VPN portals, NAS devices, monitoring tools). NCSC-NL tracks 972 products, 212 of them marked vulnerable [9]. Those boxes don’t show up when you run php -v on your servers.
  • Stopgap limits: session.upload_progress.enabled=0 removes one consumer path for CVE-2026-109619. It does nothing for CVE-2026-109618. The advisories say updating is the only complete fix [1].

2. Node2Shell / HTTP2Shell (CVE-2026-111497)

Disclosed 21 October. The bug is in Google’s Brotli decoder, which Node.js bundles. Fixed in Node.js 22.23.4, 24.22.1 and 26.12.1, and in Brotli 1.0.10, 1.1.1 and 1.2.1 [10][11].

Needs

  • Attacker-controlled Brotli data has to reach the decoder. In practice that happens through three routes:
    • Inbound request bodies sent with Content-Encoding: br, if the app uses a parser that decompresses them. Express 5 / body-parser 2 parsers do this by default. So do common Koa parsers and @fastify/compress with request decompression enabled [12][13].
    • Outbound responses. The built-in fetch() decodes Brotli responses automatically, and “Sending Accept-Encoding: identity does not prevent it” [13]. That makes webhooks, link previews and anything that fetches a URL the user supplies the realistic client-side path.
    • Direct node:zlib Brotli calls on untrusted data [10].
  • Not exposed by default: a bare node:http server (“doesn’t decode incoming Brotli bodies on its own”, [13]) and Express 4 / body-parser 1.x, which only handle gzip and deflate [13].
  • For RCE, a build-specific exploit. ELLIO needed a separate offset for each Node build, and most of its 400 Express attempts crashed the worker rather than returning a shell [14]. Imperva’s demonstration “did not succeed on every attempt” [15].

Realistic worst case

  • A worker crash for most exposed apps. RCE is possible against a known build where an attacker can iterate.

Exploitation status: attempts observed; no confirmed compromise.

  • CISA added the CVE to the KEV catalogue on 27 October [16].
  • Sensor networks have logged attempts. Previdian is careful to add: “They do not show that the target was vulnerable or that the attempt succeeded” [17].
  • ELLIO tested the attack chain it captured in the wild and got zero hits in 50 attempts [14].

Name check

  • “HTTP2Shell” has nothing to do with the HTTP/2 protocol. It’s the same CVE, renamed as more HTTP stacks turned out to bundle Brotli [18].
  • The blast radius is real, but it is conditional: .NET says “Request decompression and automatic response decompression are not enabled by default” [19], and Envoy is affected only when its Brotli decompressor is enabled [12].

Check first

  • node -v on hosts, in CI and inside every container image.
  • npm ls express body-parser koa @fastify/compress in each service.
  • pip show brotli brotlicffi in Python services.
  • Stopgap limits: rejecting Content-Encoding: br at the proxy, or setting express.json({ inflate: false }), protects inbound bodies only. It does nothing for fetch() responses, and inflate: false also rejects gzip bodies [13].

3. Java4Shell / Java2Shell (CVE-2026-114938)

Disclosed 17 November. The bug is in the HotSpot x86-64 Base64 decode intrinsic. Fixed in 17.0.21.1, 21.0.13.1, 25.0.5.1, 27.0.1.1 and the Oracle 8 performance build 8u513 [20][21].

Needs

  • Specific hardware. “Java4Shell also requires an x86-64 processor with AVX-512 VBMI support. ARM64 CPUs are not affected” [22]. That means Intel Xeon Ice Lake or newer, or AMD Zen 4 or newer [22]. Graviton and Apple Silicon are not affected.
  • Specific JDKs. 17.0.6 and later, 21, 25 and 27. Standard Java 8 and 11 and OpenJ9 are not affected [20].
  • A warm decoder. The vulnerable fast path is only used after the C2 JIT has compiled the decode routine [20].
  • The right input. Attacker-controlled Base64 has to reach java.util.Base64’s basic decoder, ending in a stray = [22]. OX says the MIME decoder and third-party Base64 libraries are not affected [21].
  • For RCE, work per target. Calif: “There is no universal payload.” Its exploits rely on heap spraying and retries, and “Low, stable round-trip times are preferred for a more reliable attack.” Calif ran the attack from a VM it rented in the same cloud region as the target [22].

Realistic worst case

  • A crash. For affected hardware and JDKs, “a short burst of requests can reliably bring an affected service down” [21].
  • RCE has been shown against default configurations of Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate [22], but each target needs its own tailored exploit.
  • OX estimates that fewer than 0.12% of JVMs are RCE-exposed and about 0.8% are crash-exposed. That figure is a chain of estimates, and OX notes its own method “likely understates risk” [21].

Exploitation status: none reported as of 18 November [21].

  • The public PoC demonstrates the memory corruption only. Calif has not released its weaponised exploits [22][23].
  • Severity is reported as CVSS 4.0 9.2 (Calif) and CVSS 3.1 8.1 (Oracle) [23].

Name check: “Java4Shell” invites comparison with Log4Shell. Log4Shell needed one string anywhere in a log line and worked on any CPU. This one needs a particular CPU, a particular decoder and a per-target exploit.

Check first

On the host or the cloud instance type:

grep -m1 '^flags' /proc/cpuinfo | tr ' ' '\n' | grep -E '^(avx512f|avx512bw|avx512vbmi|bmi2)$'

All four flags must be present [20][21]. Then check the JDK version bundled inside each product (Jira, Jenkins, Keycloak, Elasticsearch), not just the java on your PATH.

  • Stopgap limits: -XX:-UseBASE64Intrinsics disables only the vulnerable intrinsic [21]. -XX:UseAVX=2 disables all AVX-512 and costs performance elsewhere [24], and on ARM it stops the JVM from starting [20]. Both are temporary measures.

4. NGINX2Root (CVE-2026-123007 and four others)

Disclosed 18 November. Fixed in nginx 1.30.7 and 1.31.9 [25].

Needs

  • For the lead bug, CVE-2026-123007:
    • HTTP/2 enabled on the listener [26].
    • A lot of traffic. Rapid7 says RCE “requires a significant amount of network traffic to be transmitted, approximately 128 GB of HTTP/2 data” [27]. That is about two minutes at 10 Gbit/s or seventeen at 1 Gbit/s. Below roughly 300 Mbit/s, the attack outlasts nginx’s default one-hour connection lifetime.
    • A direct connection to nginx. If the HTTP/2 connection is proxied, Rapid7 says the frames will likely not be retransmitted in a usable way [27].
    • An ASLR bypass. Per F5, code execution is possible only if ASLR is disabled or bypassed [28]. But other researchers say the HTTP/2 bugs can leak the addresses needed.
  • The other four bugs each need their own configuration [25][26]:
    • CVE-2026-123633: HTTP/2, plus proxying, plus pressure on worker_connections.
    • CVE-2026-123634: a resolver with a request-derived proxy_pass (e.g. proxy_pass http://$host;), plus mirror.
    • CVE-2026-121810: HTTP/3 (QUIC), plus proxying.
    • CVE-2026-113425: the mail proxy with auth_http.
  • Impact is scoped to the worker. Xint: “The commands run with the privileges of the nginx worker process that handles the request” [26].

Realistic worst case

  • A worker crash. RCE as the nginx worker user is possible for a direct-to-origin HTTP/2 listener that an attacker can flood with traffic and probe.

Exploitation status: none reported, and not in the KEV catalogue as of 19 November [29].

  • Rapid7’s public exploit is meant only to demonstrate the technique “against a local freshly started Docker container”. It will likely fail against a server with multiple concurrent clients or a busy heap [30].
  • Censys’s quick test, which has been cited as a reproduction, confirms only the flow-control accounting error: “Accounting error confirmed over h2c on 1.31.8; 1.31.9 rejected the sequence. No crash.” [31].

Name check

  • “NGINX2Shell” is a second name for CVE-2026-123007 [32], not a separate bug. F5 and nginx.org use neither name [33].
  • One name covers five CVEs, and each needs a different configuration. Disabling HTTP/2 isn’t a fix for “NGINX2Root” as a whole.

Check first

On each instance:

nginx -V 2>&1 | head -1
nginx -T 2>/dev/null | grep -nE '\bhttp2\b|quic|http3|^\s*mail\b|\bresolver\b|\bmirror\b|proxy_pass\s+https?://\$'

This tells you which of the five cards apply to you.

  • Also check whether HTTP/2 terminates at your CDN or load balancer, or at nginx itself.
  • Stopgap limits: turning off HTTP/2 covers CVE-2026-123007 and CVE-2026-123633 only [26][34]. It breaks gRPC [34] and does nothing for the resolver, HTTP/3 or mail bugs.
  • After upgrading, restart. A reload leaves the old binary running [26].

5. Ruby2Shell (GHSA-34m4-m9fh-gqp5)

Disclosed 18–19 November, alongside six other regexp and string flaws. Fixed in Ruby 3.3.14, 3.4.13 and 4.0.10. Ruby 3.2 and older get no fix [35].

Needs

  • Untrusted text in a legacy single-byte encoding. The fix touches 14 encodings: ISO-8859-1, -2, -3, -4, -9, -10, -13, -14, -15 and -16, plus Windows-1250, -1252, -1254 and -1257 [36]. UTF-8 is not affected [37].
  • A case change on that text. The researchers name upcase, swapcase, downcase(:fold) and casecmp? [37]. The bug comes from ß, which becomes the two characters “SS” when upcased, while the loop only checked for one byte of room [36].
  • A narrow primitive. The overflow is one byte, and that byte is always S (0x53) or s (0x73) [37]. An attacker can influence where it lands, but not what gets written.
  • The framework path. Veria Labs lists Rails and classic-style Sinatra as affected “in their default configuration”, and plain Rack as not exposed by default. It has not published which input reaches the bug [37].

Realistic worst case: still unknown, and this card is the one most likely to change.

  • The Ruby team says: “We are aware of PoCs that prove this flaw can be used to trigger RCE in some scenarios” [38].
  • Veria Labs: “We’ll be publishing a proof of concept in a week, once people have had time to upgrade” [37].

Exploitation status: none reported. There is no public PoC yet (as of 19 November 2026).

Name check

  • Much of the coverage quotes CVE-2026-132825 with a CVSS 3.1 score of 7.4 [38].
  • If your scanner keys on the CVE ID, check that it also matches the GHSA ID.

Check first

  • Run ruby -v in the app’s own environment. For running workers, use readlink -f /proc/<PID>/exe.
  • Check the bundled Ruby in self-hosted products as well. Veria lists GitLab, GitHub Enterprise Server and Chatwoot releases as affected [37].
  • Search the code for force_encoding or encode into the encodings above, followed by case changes.
  • Stopgap limits: there is no official workaround [35]. Converting text to UTF-8 before changing case protects your own code, but not the undisclosed framework path.

The Pattern: Volume Isn’t Evidence

Put the five cards next to each other:

HardwareConfig / code pathNetwork positionPublic RCE exploitIn the wild
PHP2ShellAnymod_php direct; FPM via app session/unserialize() op; upload progress on by defaultExposed portals / appliancesYes (vendor/researcher)Yes, KEV 10-26
Node2ShellAnyBrotli-decoding body parser, fetch()Any client of attacker dataDemonstrated, unreliableAttempts, KEV 10-27
Java4ShellAVX-512 VBMIBasic decoder, trailing =, warm JITLow RTT preferredNo (crash PoC only)No
NGINX2RootAnyHTTP/2 (123007); others per moduleDirect to origin, ~128 GBFresh-container onlyNo
Ruby2ShellAnyLegacy encoding + case changeUndisclosedNo (PoC in ~1 week)No

Two things stand out.

The real impact is not well understood yet. The three November bugs got the biggest headlines, but none of them has a public RCE exploit that works outside a lab or any observed exploitation as of 19 November 2026. The two October bugs are the ones on the KEV list.

The prerequisites move. Node2Shell was described as “DoS only” in its first days, and independent RCE demonstrations followed within a week [13][14][15]. Ruby2Shell’s PoC is due in about a week. A card that says “crash only” today is a reason to schedule the patch, not to skip it.

Where a Specialist Helps

None of this replaces patching. It decides the order of patching and the urgency of each item, and that depends on facts only your environment can answer. This is the work we do with customers at Conviso when a wave like this one lands:

  • Inventory beyond the obvious. This means the runtimes bundled inside appliances, vendor products and container base images, not just the php -v or java -version on your own servers. In two of the five cards, that’s where the exposure was. A runtime installed from OS packages or a base image often won’t appear in a dependency manifest, which is why our PHP2Shell analysis warned that SCA alone isn’t enough. In the Conviso Platform, Vuln Intelligence brings findings from Conviso and third-party scanners into one queue, ranked by business risk.
  • Reachability. Does attacker input actually reach the vulnerable code path? That covers the decoder, the parser, the encoding and the nginx module. It’s a code and configuration review question, not a version check. AI Secure Code (SAST, SCA, secrets and IaC) covers the code side. Architecture review and threat modeling cover how requests flow through your stack.
  • Environmental conditions. CPU generation per instance type. Whether HTTP/2 terminates at a CDN or at the origin. Which SAPI is in use. Whether management interfaces are reachable from the internet. When the answer is uncertain, a targeted pentest settles it in your environment, not in a lab. The Offensive & Audit Manager tracks it from scoping to retest.
  • Compensating controls, and what they don’t cover. Each card’s stopgap has limits. An edge WAF such as Conviso’s Web and API Protection can reduce exposure while you patch, but it doesn’t replace the update. It also doesn’t help with paths that never cross the edge, such as fetch() responses or a management portal reached from inside the network.
  • Tracking exploit maturity, so a “schedule it” item becomes “now” the day a working PoC appears. The Conviso Research Team follows these disclosures as they evolve. The Ruby2Shell card above is the next one to watch.

The result is a short, ranked list for your business, instead of five alarming headlines: patch now, patch this cycle, monitor. Some of these vulnerabilities will be urgent for some companies and routine for others. Getting that call right protects the business without wasting effort or causing panic.

If you want help working through that list for your own applications and infrastructure, talk to Conviso.

Sources

  1. PHP GHSA-w8hw-w9fx-j79w Primary source
  2. PHP GHSA-6wmf-4rv5-p3rp Primary source
  3. Calif, PHP2Shell Primary source
  4. Tenable PHP2Shell FAQ Secondary source
  5. CISA KEV catalogue Primary source
  6. The Register Secondary source
  7. Unit 42 Primary source
  8. Check Point Research Primary source
  9. NCSC-NL php2shell product list Secondary source
  10. Node.js October 2026 security releases Primary source
  11. google/brotli releases Primary source
  12. Cycode Secondary source
  13. Sysdig Secondary source
  14. ELLIO Secondary source
  15. Imperva Secondary source
  16. CISA alert, 27 Oct 2026 Primary source
  17. Previdian Secondary source
  18. SecurityWeek, Node2Shell exploitation Secondary source
  19. .NET out-of-band updates Primary source
  20. JFrog Secondary source
  21. OX Security, Java4Shell Secondary source
  22. Calif, Java4Shell Primary source
  23. BleepingComputer, Java4Shell Secondary source
  24. Sweet Security Secondary source
  25. nginx security advisories Primary source
  26. Xint, NGINX2Root Primary source
  27. Rapid7, CVE-2026-123007 Primary source
  28. SecurityWeek, NGINX2Root Secondary source
  29. watchTowr, NGINX2Root FAQ Secondary source
  30. Rapid7 PoC Primary source
  31. Censys Secondary source
  32. Crypto Briefing Secondary source
  33. Tenable, NGINX2Root FAQ Secondary source
  34. Qualys Secondary source
  35. ruby-lang.org advisory Primary source
  36. Ruby fix commit Primary source
  37. Veria Labs, Ruby2Shell Primary source
  38. Ruby GHSA-34m4-m9fh-gqp5 Primary source
  39. Trail of Bits on X Primary source
  40. Cloudflare, PHP CTF Secondary source
  41. PHP2Shell Explained: Critical PHP RCE Vulnerabilities, Impact, and Remediation Previous post, Conviso Research Team, 22 Oct 2026

Próximo passo

Mais recentes

Ver mais do blog