These are nine Apache case files, each about one bug. Stacked together, they write the same sentence. The vendor keeps shipping security as a line every author has to remember, and every fix is the same invocation, copied into one more file.
The Nefarious Plan of Apache Software Foundation
The comment named one consumer. The CVE names three more. The grep on trunk this morning names four that the CVE still does not.
That is the last line of CVE-2026-24072: AP_EXPR_FLAG_RESTRICTED Was Opt-In For Fifteen Years. The Patch Opts In Three Modules. Four More Still Read /etc/passwd From .htaccess.. The flag has existed since 2011. Two modules in the whole tree ever set it before 2026. The CVE made that number five. The grep on trunk still finds four callers that forgot.
Nine case files sit behind this one. All on Apache Software Foundation. All written alone. Stack them and the same sentence writes itself. Apache ships security as a line each author has to remember. The mechanism is there. The invocation is not. The fix, in every file, is the invocation. Copied. Into one more place.
The fix is always a line somebody else already wrote
CVE-2026-40858: Same Five Lines. Same Author. Eight Apache Camel Components. is the clean case.
The CVE description names three. The git log names eight.
Andrea Cosentino signed the camel-sql ObjectInputFilter on January 9, 2024, and signed the camel-infinispan one on April 14, 2026. Same five lines. Twenty-seven months apart. Six other components got the same patch in between. The deserialization substrate, DefaultExchangeHolder implements Serializable plus ObjectInputStream.readObject(), has not moved. No module in camel-support owns the deserialization step. Each component rolls its own.
CVE-2026-40453: Camel Patched One Header Filter In 2025. Five More Needed The Same Line. is the same shape in the same vendor, one class higher up the tree.
DefaultHeaderFilterStrategy.lowerCasestill defaults tofalsein master. The next subclass that extends it and is used to filter inbound headers will inherit the same one-line patch waiting to be remembered.
One line, setLowerCase(true);, added to HttpHeaderFilterStrategy in March 2025 to close CVE-2025-27636. Thirteen months later, the same author lands the same line in five more subclasses to close CVE-2026-40453. The base class default has not changed. The next subclass inherits the hazard, not the fix.
The ap_expr file is the oldest substrate of the nine.
The fix is not a new mechanism. The mechanism has existed for fifteen years. The fix is three lines that say please use it, in three of the modules that never did.
Three different Apache products. Three different files. Same sentence.
Opt-in defaults are the plan
The NiFi file is the one that puts the plan in writing. CVE-2026-39816: Apache NiFi Patched One Annotation. Ten Days Later, Apache Removed the Mechanism. describes a class annotation every code-executing component author was supposed to remember to write at the top of their class.
EXECUTE_CODE is whichever components have written
@Restricted(requiredPermission = RequiredPermission.EXECUTE_CODE)at the top of the class. Forget the tag and the framework has nothing else to consult.
Sixteen canonical components wrote it. Three code-executing classes in the graph bundle did not. The CVE is the gap. Ten days after the patch, Apache merged NIFI-15847 and deleted the annotation from sixty-one classes. The NiFi Improvement Proposal that governs the deletion is titled "Removal of Restricted Component Authorization." The release that carries the fix for CVE-2026-39816 is the last release in which the mechanism the fix uses will exist. That is not me interpreting. That is the NIP's own title.
The file then says the sentence the other eight files are repeating in different words.
Authorization based on a class-level annotation that every component author has to remember to write is an opt-in security boundary, and one missed opt-in is one CVE.
ActiveMQ repeats it with a Set.of(...) literal. Camel repeats it with setLowerCase(true). HTTP Server repeats it with a parse flag that is zero by default. mod_http2 repeats it with an assertion the producer side of the same file is not enforcing.
Two halves of the same concept, written separately, and nobody read both
CVE-2026-25604: The PoC Tests /login. The Bug Is on /login_callback. is the same disease at the level of one file.
Two fields. Same call. Same concept, what is this server's hostname? One hardcoded to configuration. One derived from whatever the client sent.
_init_saml_auth builds the SAML ACS URL from conf.get("api", "base_url"). _prepare_request builds http_host from the inbound Host header. Both get passed to OneLogin_Saml2_Auth(request_data, merged_settings) in the same call. Same concept, two sources, one contradiction, no reviewer caught it for a year. The fix is four lines. The failure mode is reading one function and not the next.
CVE-2026-23918: m->spurge Was an h2_ihash. The Array That Replaced It Kept the Assertion, Not the Dedup. is the same shape in mod_http2 over fifty-two months.
The producer changed from "container guarantees uniqueness" to "caller is trusted to not double-add." The consumer, which had previously relied on the same container guarantee, was rewritten to scan the array for a duplicate and abort if it found one. Producer enforcement was deleted. Consumer enforcement was preserved as a panic.
An h2_ihash was swapped for an apr_array. The consumer-side assertion ap_assert("stream should not be in spurge" == NULL) kept the invariant's name. The producer-side guarantee, which was the hash table's deduplication, was deleted. For fifty-two months three producers pushed onto that array without checking, and the consumer aborted the process when it noticed.
The assertion was telling the truth. The truth was that nobody was enforcing the invariant anymore. Nobody read both halves of the same file.
The branch is also an author
CVE-2026-42779: Apache MINA Shipped the Fix on One Branch. The Same Day, Two Other Branches Shipped the Bug. is where the opt-in sits on the release engineer.
The fix was in the codebase on 2026-04-24. Two branches shipped that day without it. The window was five days, and the commit message labeled the fix a backport of itself.
On 2026-04-23, the maintainer committed the fix to 2.0.X with the message "Added patch regarding whitelist bypassing backported from mina-2.2.X." There was no 2.2.X commit. He was labeling his own fix a backport of a commit that did not exist. On 2026-04-24, three releases shipped. One carried the patch. Two did not. The fix landed on 2.1.X and 2.2.X five days later with the commit message "Added a missing fix."
Three live branches. Three release cuts. Each branch is a person who had to remember. One of them did.
The deny list is the admission
CVE-2026-34197: ActiveMQ's 5.19.4 Patch Denied One Scheme. Six Weeks Later, the Same Line Denies Fourteen. and CVE-2026-40466: HTTP Discovery Was Row Two Of A Fourteen-Row Deny List are the plan written as a Set literal.
The CVE record names one scheme. The deny list ends in
masterslaveuntil the next commit decides it does not.
5.19.4 banned the vm scheme. Six weeks later the same file bans fourteen. One of those fourteen got a CVE because a researcher wrote a Jolokia POST and reported it. The other twelve got patches with PR titles ending in "part 2." The maintainer knew. The codebase was being read in waves by its own author. Only the rows scanners noticed got numbers.
The deny list has fourteen entries. Two have CVEs.
A deny list is an opt-in security boundary with a keyboard instead of an annotation. Each row is a name the author had to remember. The parameter type is still String. The URI registry is still open. The signature of addNetworkConnector(String) has not changed since 4.x.
Where you are right
The engine is not the charge. Java's ObjectInputStream.readObject() is a known hazard the language ships with. AP_EXPR_FLAG_RESTRICTED is a correct mechanism and does what the comment above it says. The @Restricted annotation enforces what the NiFi authorization layer reads out of it. _init_saml_auth in Airflow hardcodes the ACS URL correctly; the PoC's mock is a stronger version of the bug than the real code. Stefan Eissing's assertion in h2_mplx.c was named with the right words; the invariant it checks is real. Christopher Shannon's validateAllowedScheme is correct code; it does what its body says.
None of the controls in these files are broken. Every one of them does its job.
The charge is the shape of the ship. Apache keeps publishing correct mechanisms as idioms every author has to invoke. The call site is the gate. The gate is in every author's head.
Where the next one comes from
Open BrokerView.java on activemq-5.19.x. Read the DENIED_TRANSPORT_SCHEMES literal. The transport factories registered under META-INF/services/org/apache/activemq/transport/ that are not on that list are the next rows. A PR title ending in "part 3" is the signal.
Open include/ap_expr.h on httpd trunk. Read AP_EXPR_FLAG_RESTRICTED. Grep for ap_expr_parse_cmd(. Read the override class on every containing directive. Any directive declared OR_FILEINFO or OR_OPTIONS whose parse site does not pass the flag is a candidate. The ap_expr file names four of them: mod_headers, mod_alias, mod_substitute, mod_filter.
Open camel-support. Search for a shared deserialization codec. There is not one. Each Camel component's readObject() is its own site. The ninth is whichever component adds remote persistence next.
Open NIP-24. NiFi 2.10.0 removes @Restricted entirely and shifts the gate to the standard component-edit policy. Audit which NARs an operator can load and who has component-create rights, because that is the only gate 2.10.0 will have.
None of this is an exploit. It is the directory listing of what the vendor has written down and left opt-in.
What the next advisory will say
The next Apache advisory will do one of three things.
A row will be added to DENIED_TRANSPORT_SCHEMES in BrokerView.java on activemq-5.19.x, with a PR title naming a transport scheme already present in META-INF/services/org/apache/activemq/transport/.
Or AP_EXPR_FLAG_RESTRICTED will be added to one of mod_headers.c, mod_alias.c, mod_substitute.c, or mod_filter.c, in a commit whose diff is three lines with a cmd->pool == cmd->temp_pool check already written by r1933350.
Or a ninth Camel component will get its own copy of ObjectInputFilter.Config.getSerialFilter() in a commit signed by Andrea Cosentino.
The prediction is called when any one of those three lands. The prediction is missed if the next Apache advisory does none of them.
To Apache. The plan is in your commit log. You wrote "part 2" in a commit title, labeled a fix the backport of itself, and left "e.g." standing for fifteen years, and the fix for the next one is the invocation you already wrote, copied into one more file.