[ <- back-issue archive · live wire · issue #245 · HUB 24H · 2026-11-20 17:53 UTC ]
 .-.                       .-. 
(: `-----------------------' :)
 )   issue #245 // hub 24h   ( 
(: .-----------------------. :)
 `-'                       `-' 
> /var/archive/digest/245.bone :: [ HUB 24H // 53 sources analyzed // 2026-11-20 17:53 UTC ]

Wrong-Math Week: Unsigned Guards, Inverted Flags, and Band Edges That Fall Through

TL;DR: Nearly every hard hit on the wire this cycle traces to a check written in the wrong arithmetic, from PHP's non-atomic session writes and ytnef's unsigned INT_MAX guard to pam-mysql's inverted PAM flag and a HackRF band-edge comparison that falls through; meanwhile Google's $10K auto-disclosure rule rewrites bounty economics and KrakenSDR learns to home 121.5 / 243 MHz distress beacons.

The interpreter and parser layer is bleeding from a single wound: bounds and flag checks evaluated in the wrong arithmetic. PHP's 17-fix drop across 8.2.36 through 8.6.0 leads with 'phpAfterShell' (CVE-2026-126714), where the default session.save_handler = files path writes session update progress non-atomically, so a killed worker leaves a half-written session file that the next session_start() deserializes straight into an object injection primitive, gadget chain permitting; its sibling CVE-2026-129289 smuggles a hidden FastCGI request through leftover unread body bytes on persistent backend connections. The identical wrong-math shape resurfaced in ytnef 2.1.3: the INT_MAX guard on the stream-supplied uncompressedSize, bolted on in 2017 for CVE-2017-6302, is evaluated in unsigned arithmetic, so the allocation wraps and the 207-byte RTF prebuffer overruns the heap by 1 to 207 attacker-chosen bytes, reachable by any email sender through ytnefprocess calendar items in ytnef/vcal.c with zero user action; Evolution, Geary, and the Claws TNEF plugin all chew winmail.dat. Ruby matched with CVE-2026-132825, a one-byte heap overflow in string case conversion where upstream admits PoCs prove RCE in some scenarios. One byte. PoCs exist. And @sigdevel stacked three MP4Box bugs into a single GPAC fix commit 5439b94: cyclic DASH dependencies recurse dasher_gather_deps() past filters/dasher.c:3777 to stack exhaustion (CVE-2026-130635), an overlong NHML filename writes 1 byte past the GF_MAX_PATH stack buffer at filter_pid.c:9366 (CVE-2026-130793), and an unchecked strcat() in gf_mpd_resolve_url() writes 37 bytes past a 110-byte heap region (CVE-2026-130634), per the advisory. glibc even joined in with CVE-2026-126765, a glob double free on the allocation-failure path where rewritten names are freed but their pointers stay in the result vector for globfree to hit again. Error paths and unsigned math: twenty years of the same lesson.

The auth-bypass column is the same disease wearing a different coat. pam-mysql (CVE-2026-124749, CVSS 9.8) inverts PAM_DISALLOW_NULL_AUTHTOK at both pam_sm_authenticate call sites, passing !(flags & PAM_DISALLOW_NULL_AUTHTOK) where it should pass the flag straight, and since the password compare is nested inside `if (row0] != NULL)`, any account whose stored password column is SQL NULL accepts any password, pre-auth, on a network socket. The sting is architectural: stock sshd with `PermitEmptyPasswords no` is the vulnerable config, so the administrator who did everything right gets the permissive path. Java4Shell ([CVE-2026-114938) is arithmetic underflow at silicon speed: a malformed = padded tail pushes the calculated output length below zero in the OpenJDK Base64 intrinsic, the AVX-512 write mask goes all-bits-set, and 64 bytes land where at most two belong, 62 past the buffer; it needs 17.0.6+, AVX-512 VBMI, and roughly 10,000 decodes to JIT, and -XX:-UseBASE64Intrinsics kills it. The nginx HTTP/2 RCE pair CVE-2026-123007 and CVE-2026-123633 has a messier tail: per the distro thread, Debian trixie is fixed at 1.26.3-3+deb13u11 but bookworm is still naked, RHEL 8/9/10 are listed affected, and CentOS 7/8 plus Debian 8-11 get nothing but a reporter's unofficial backports. On the researcher wire, @ctbbpodcast only teased "We live for chains like this!" without naming a target, a CVE, or a primitive, so the chain stays on the cutting-room floor until they publish; treat it as a marker that somebody's write-up drops soon, not as intel you can act on. Upload validation is theater when the storage layer gets a vote in the headers.

Disclosure economics shifted the same week, and hunters should read the fine print before December. Google VRP announced that any report rewarded at $10K or higher is automatically published on bughunters.google.com from the start of December, no negotiation, no quiet burial; @erbbysam's read is the right one, go be the researcher whose reports you'd actually want to read, and @brutecat's two-word verdict ('transparency is king') says the quiet part loud. Your chain-of-discovery narrative, reproduction steps, and payload math are about to become permanent public artifacts with your name on them, so write reports like documentation, not tickets. Pair that with the one repo worth cloning this stretch: vitorfhc/videt, a Python monitor that watches major open-source trees for silent security patches in the 'negative-day' window before a CVE is assigned and then checks whether the fix is complete or still bypassable. That tool exists because this week's pattern is structural: the ytnef maintainer shipped 2.1.3 with a self-contained fix in lib/ytnef.c and fresh regression tests in test-data/test_decompress_rtf.c, but the CVE for the heap write is still pending, which means every unpatched downstream consumer is living in exactly the window videt is built to flag. Point it at sibling TNEF and LZFu decompressors, diff pam_ldap and pam_krb5 for the same inverted-flag shape, and grep other SIMD Base64 decode paths for = tails that push computed length negative.

The RF desk closed the loop on the boundary-bug theme with hardware evidence. KrakenSDR Suite V2 is now pure C++ with Heimdall holding all five tuners sample- and phase-coherent for live MUSIC direction finding, and beacon121 extends that coherence into swept-tone ELT/EPIRB/PLB homing at 121.5 / 243 MHz with per-beacon bearings: a real COSPAS-SARSAT homing rig on a $500 coherent board, and a ready-made calibration source if you want to verify phase coherence before you trust the bearings. No HackRF firmware news crossed the wire this cycle, so if you were holding out for a band-edge tuning fix, keep holding; the only SDR movement worth a diff was KrakenRF's coherence work above, and the band-seam sweep stays on your own todo list either way. Over in 802.11 land, morrownr/mt76 quietly rolled MT7925 WiFi firmware back to 20260813 with no stated reason, and an undocumented rollback on a monitor-mode-capable driver is a diff you open first: either the newer blob broke injection for somebody in the field, or it broke something quieter. ESP32 Marauder merged PR #1701 with framed SD file streaming over serial, so a deployed wardrive board empties its capture card over the wire; that framing protocol is a fresh parser on a hostile port, so feed it malformed frame lengths before you trust it with your only copy of the evidence. And if you operate up north, the RNT Foundation's read of the CCA report is sobering: Canada runs zero operational VLBI or SLR stations and only sparse GNSS Active Control Stations across the Arctic Archipelago, so anyone spoofing GNSS up there is working against a degraded verification net. Wrong arithmetic at the top of the stack, thin reference frames at the bottom; hunt the primitives, not the press releases.

 .-.                     .-. 
(: `---------------------' :)
 )   intel brief // #245   ( 
(: .---------------------. :)
 `-'                     `-' 
> /var/archive/digest/245.intel :: [ 2026-11-20 17:53 UTC ]
  • ytnef 2.1.3 heap overflow in DecompressRTF: The INT_MAX guard added for CVE-2017-6302 is evaluated in unsigned math, so a stream-supplied uncompressedSize slips past it, the allocation wraps, and the 207-byte RTF prebuffer overruns the heap by 1 to 207 attacker-chosen bytes. ytnefprocess reaches it via calendar items in ytnef/vcal.c on mail delivery with zero user action; Evolution, Geary, and the Claws TNEF plugin all chew winmail.dat [Yeraze/ytnef, oss-security].
  • pam-mysql pre-auth sshd bypass: Both pam_sm_authenticate call sites at pam_mysql.c:4708 and :4799 pass !(flags & PAM_DISALLOW_NULL_AUTHTOK), which is 0 exactly when the app set the flag, so vresult = null_inhibited returns PAM_MYSQL_ERR_SUCCESS. Password compare is nested inside if (row[0] != NULL), so any account with a SQL NULL password column accepts any password. The sting: stock sshd with PermitEmptyPasswords no is the vulnerable config, CVSS 9.8 [oss-security].
  • phpAfterShell session file object injection: Non-atomic session writes mean a killed worker leaves a malformed session file that session_start() happily deserializes into an object injection primitive; gadget chain permitting, that is unauthenticated RCE with session.save_handler = files. CVE-2026-129289 smuggles a hidden FastCGI request through leftover body bytes on persistent backend connections. Seven of the 17 fixes are CVSS 9.2 [Wiz Research].
  • Java4Shell AVX-512 Base64 heap overflow: A malformed = padded tail underflows the calculated output length below zero in the OpenJDK Base64 intrinsic, the write mask goes all-bits-set, and 64 bytes land where at most two belong, 62 past the buffer. Needs 17.0.6+, AVX-512 VBMI, and roughly 10,000 decodes to JIT; -XX:-UseBASE64Intrinsics kills it [Calif via Hackaday].
  • Flatpak sandbox bypass CVE-2026-34078: Codean Labs calls it a complete sandbox escape out of their GNOME evaluation, rooted in the finish-args permission model and XDG Desktop Portal surface. Worth reading before you trust any --share=network app manifest [Codean Labs].
  • Ruby one-byte heap overflow in case conversion: GHSA-34m4-m9fh-gqp5, CVE-2026-132825: changing the case of a crafted string overruns the heap by one byte, and upstream admits PoCs prove RCE in some scenarios. One byte is plenty when the heap layout cooperates [ruby/ruby via oss-security].
  • GPAC dasher recursion and template overflow: @sigdevel stacked three MP4Box bugs: cyclic DASH dependencies recurse dasher_gather_deps() to stack exhaustion at filters/dasher.c:3777 (CVE-2026-130635), an overlong NHML filename blows the GF_MAX_PATH stack buffer at filter_pid.c:9366 (CVE-2026-130793), and gf_mpd_resolve_url() unchecked strcat() writes 37 bytes past a 110-byte heap region (CVE-2026-130634). All fixed in 26.07.0 commit 5439b94 [oss-security].
  • glibc glob double free: CVE-2026-126765 in 2.44 and earlier: on allocation failure while glob prepends the directory prefix, rewritten names are freed but their pointers stay in the result vector, and globfree frees them again. Only reachable near memory limits under RLIMIT_AS or on 32-bit, but that is exactly where long-running daemons live [glibc, oss-security].
 .-.                       .-. 
(: `-----------------------' :)
 )   possibilities // #245   ( 
(: .-----------------------. :)
 `-'                       `-' 
> /var/archive/digest/245.possibilities :: [ 53 sources ]
  • Hunt sibling TNEF parsers for the same unsigned guard: The unsigned INT_MAX check pattern is a classic copy-paste slip; grep other LZFu and RTF decompressors (Outlook-adjacent libs, ripMIME, amavisd helpers) for size guards done in unsigned math, then fuzz calendar-item TNEF attachments through ytnefprocess on unpatched 2.1.2 [inspired by Yeraze/ytnef].
  • Audit every PAM module for inverted flag logic: The pam-mysql bug is a boolean inversion at two call sites with the sibling chauthtok sites passing a literal 0, which is what exposed the slip; diff pam_ldap, pam_krb5, and pam_pgsql for the same !(flags & PAM_DISALLOW_NULL_AUTHTOK) pattern and test with NULL password columns against stock sshd [inspired by the pam-mysql advisory].
  • Weaponize the PHP session file primitive on shared hosts: Non-atomic session writes plus session_start() deserialization is a gift on any multi-worker PHP-FPM box; race a kill against a large $_SESSION write, then chain POP gadgets from the target's composer deps. Also probe FastCGI persistent connections for leftover-body smuggling per CVE-2026-129289 [inspired by Wiz Research].
  • Fuzz Base64 tails on other SIMD intrinsics: The underflow-to-all-bits-mask bug class is not Java-specific; check Go, Rust, and native OpenSSL AVX-512 decode paths for malformed = padded tails that push computed output length negative, and remember the ~10,000-decode JIT warmup when building harnesses [inspired by the Java4Shell writeup].
  • Chase the remaining six Convert::ASN1 CVEs: Only CVE-2026-126540 (quadratic os2ip bigint loop) and CVE-2026-126568 (inverted $tb || guard plus MSG_PEEK buffer doubling) have mechanics published; pull the 0.37 diff and reverse the other six from the patch hunks, then hammer LDAP-dependent Perl apps with oversized INTEGERs [inspired by CPAN Security Group].
  • Diff the GPAC fix commit for incomplete patches: One commit 5439b94 claims to fix recursion, stack overflow, and heap overflow across three subsystems; grep sibling template-expansion call sites for other unchecked strcat() and GF_MAX_PATH copies, and check whether the cycle detection in dasher_gather_deps() covers indirect dependency loops [inspired by @sigdevel].
  • Probe nginx HTTP/2 RCE patch coverage: Debian bookworm and CentOS 7/8 are still naked for CVE-2026-123633 and CVE-2026-123007; fingerprint nginx builds behind HTTP/2 front doors, and if you run those branches, pull the reporter's unofficial backports at HOSTED-POWER/nginx2root-patches before someone else does [inspired by the oss-security thread].
 .-.                      .-. 
(: `----------------------' :)
 )   github brief // #245   ( 
(: .----------------------. :)
 `-'                      `-' 
> /var/archive/digest/245.github :: [ code & repo tradecraft ]

TL;DR: The wire's code corner is thin on new offensive tooling, but videt's negative-day commit monitor is the one repo worth cloning this week, since the negative-day window is exactly where defenders lose ground before a CVE ever lands.

The patch flow this stretch was quiet, with no fresh advisories or exploit drops landing in the feed; that lull is exactly the window a commit-watching tool is built for. Meanwhile the repo pushes this stretch were mostly wordlist syncs and PoC-builder polish, so the real hunting signal stays on the advisories, not the commits.

  • vitorfhc/videt: Python monitor that watches major open-source repos for silent security patches in the 'negative-day' window before a CVE is assigned, then checks whether the fix is complete or still bypassable. Point it at any major tree and see what it flags retroactively.
  • danielmiessler/SecLists: The 74k-star wordlist collection only saw automated GitHub Action refreshes of the trickest wordlists; no new payload content, but the fuzzing lists remain the baseline grab for any session or smuggling probing.
  • JorianWoltjer/responder: Shareable HTML/JS PoC builder with custom headers via query parameters; fresh commits add an 'open in cyberchef' shortcut and nested autocomplete in SVG/XML. Handy for shipping a clean session-injection demo to a triage team.
  • caido/rswot: Rust student-status validator against a crowdsourced domain list; v0.1.144 and repeated chore: sync domains from JetBrains/swot commits are pure list maintenance, nothing offensive here.