Full motion briefing transcript

OCT 2026 · Scene 1

October incidents, in date order.

Unrelated failures that kept landing on the same kind of weakness. Tap through them in order.

• The initial 15 cases plus later broad reviews; these are not a worldwide victim count
• 21:35 JST report (Japanese) → · Read by date →
• Sound is on — toggle it top-right

15 incidents · 1 month

Repository

08.11 · Scene 2

Seventy quiet days.

Seventy days from report to release. The dangerous window was not after publication — it was just before it.

• 08-01 Calif finds several RCE paths / 08-11 reported to PHP
• 10-20 details leak; watchTowr confirms RCE
• 10-21 07:43 UTC attack attempts observed before the fix (Cloudflare)
• 10-21 14:00 fixed releases — 8.5.12 / 8.4.27 / 8.3.36 / 8.2.35

70 days from report to public

docs/00 timeline / php.net

10.14 · Scene 3

A plaintext admin key opened nine regions.

The second incident since August. The August rotation did not cover the credentials used this time.

• 10-14 01:36 entry via internal AWS keys / 01:51 system secrets
• from 12:39 all service environment variables; commands ran in 865 database services
• root cause: plaintext, non-expiring admin credentials and a publicly reachable admin plane
• customers: rotate every environment variable again, even if you did in August

865 database services touched

Zeabur report (2026-10-19)

10.15 · Scene 4

For eleven hours, the official download was poison.

A plugin's update channel is an attack path. CVSS 9.8.

• 3.0.123 downloaded between 10-15 20:02 and 10-16 07:12 UTC
• CVE-2026-111862 / CVSS 9.8 / fixed in 3.0.124
• updating is not remediation — hunt for admin accounts, extra files, scheduled tasks
• Wordfence shipped an emergency firewall rule early

9.8 CVSS · CVE-2026-111862

Wordfence PSA

10.16 · Scene 18

Six shortcuts to root, in the kernel's own core.

Hyunwoo Kim published 'Infinity Core' on oss-security: five of the six take an unprivileged user to root, and the same five escape a container to the host.

• timer, signal, mm, futex, scheduler, SLAB — no networking, no BPF, no user namespaces needed
• NVD scores four at 7.8 and two not at all; all six sit at Received and none is in KEV
• The fix already shipped: 6.6.158 / 6.12.112 / 6.18.54 / 7.2.8, scheduler in 7.2.9 — every current stable line is above it
• What works is the reboot. A patched package next to a running old kernel leaves it open

6 : 5 CVEs : root PoCs

V4bel/InfinityCore / oss-security 2026-10-16 / linux-cve-announce (98260) / docs/20

10.17 · Scene 5

Keys in a database were the door to thousands of sites.

Connection keys work as validly signed commands, so the previous security fix did not stop them.

• customer DB read: contact data, hashed passwords, 2FA settings, three kinds of connection key
• UpdraftCentral / UpdraftVault / UpdraftClone keys used to upload, activate and delete plugins
• the signatures were real, so sites with CVE-2026-10795 patched still accepted them
• UpdraftPlus 1.26.13 / 2.26.13 auto-disconnect anything connected before 10-18 22:00 BST

3 key types abused

TeamUpdraft notice / Wordfence PSA

10.17 · Scene 6

The keys were not stolen. They were too weak to be secret.

P-384 private keys that were small integers — 4, 56, 66, 115, 155, 156, 230, 237. Between 3 and 8 bits of entropy.

• eight intermediate CAs (six TLS, two code signing). A 2025-04-30 ceremony for 24 subordinates produced a second, weak key set through a wrap/unwrap flaw
• 10-17 18:05 UTC report → 19:45 Chrome blocks by SPKI → 22:59 all Firefox. Apple blocked the six TLS, Microsoft the two code-signing
• Browsers only cleaned up their own copy: Docker, Kubernetes, IoT and Python bundles remain. Go shipped CVE-2026-111943 in 1.27.3 / 1.26.10
• The sharpest near-term risk is VPN clients (Talos, Ars): “bigger cleanup than Heartbleed”, per Go's crypto maintainer

4 one private key, literally

Google (Chrome Root Program) / Scott Helme / Cisco Talos / Ars Technica / Mozilla bug 2083449 / Tailscale TS-2026-013

10.17 · Scene 22

The attempts began four days before the advisory.

AWS honeypot telemetry holds an exploitation attempt dated 2026-10-17, and working PoCs were seen twelve hours before the embargo lifted. Disclosure day was not day one.

• 482,342 requests / 122,311 sessions / 21,998 sites / 103 countries in the first 24 hours (Imperva)
• Those are attempts, not compromises: Sophos saw the Mirai install fail, and Dataminr reports no known compromises
• Still not in KEV. EPSS 5.31% / 5.76%
• Hunt from 10-17. Starting at the disclosure date loses the first four days

4 days first attempt → advisory

AWS · MadPot telemetry / Imperva · first 24 hours / docs/01 / docs/22 §3

10.19 · Scene 7

For 94 minutes, the official binary was something else.

Source and build chain were fine. What broke was the assumption that release assets cannot be replaced.

• assets replaced 10-19 between 14:47 and 16:21 UTC+2
• releases v0.11.0–v0.15.0; only downloads inside that window are affected
• root cause: the repository was not configured for release immutability
• the fix is a setting plus verification: immutable releases and signature checks

94 minutes of exposure

GHSA-862f-55f5-vmvw

10.20 · Scene 8

55 fixes, highest 7.5 — and the version column is the trap.

“Affected” and “fixed” are different columns. 21.0.12.1 is affected; the fix is 21.0.13.

• 55 CVEs / 49 exploitable remotely without auth / highest CVSS 7.5 / zero in KEV
• affected: 8u503, 11.0.32.1, 17.0.20.1, 21.0.12.1, 25.0.4.1, 26.0.2, 27
• fixed on 10-20: 8u511 / 11.0.33 / 17.0.21 / 21.0.13 / 25.0.5 / 27.0.1
• JDK 26 gets no fix (there is no 26.0.3) — move to 27

55 CVEs · 49 remote, no auth

Oracle CPU 2026-10 / IPA alert

10.20 · Scene 9

A 4.3 SSRF is the front door to code execution.

No individual score is remarkable. CISA's KEV entry says the four chain together.

• added to KEV 10-20 / due 10-23 for 96656 and 96654
• chain: any X-Plex-Token → SSRF → plugin call → admin file write → .so executed at startup
• 96651 is not in KEV, yet it reads PlexOnlineToken — account takeover
• fixed in 1.43.3.10861, but the current release is 1.43.5.11029; NAS stores lag

4/5 CVEs in KEV · due 10-23

CISA CSAF VA-26-266-01 / CISA KEV

10.20 · Scene 10

Three “Important” bugs, one outcome: another user's data.

Tomcat's three differ in mechanism and share a result. httpd's “low” is scored 7.5. Severity is not one number.

• Tomcat: auth bypass in JNDIRealm (non-GSSAPI SASL) plus HTTP/2 and AJP response mix-ups — 11.0.27 / 10.1.61 / 9.0.123
• httpd 2.4.69: mod_rewrite UAF and a mod_proxy_html out-of-bounds write. Apache says low; NVD says 7.5
• Traffic Server: two request-smuggling CVEs (8.9 / 8.7)
• Nothing in KEV — no published exploitation is not the same as safe

9.1 Tomcat auth bypass · CVSS 4.0

Apache Tomcat security / Apache httpd 2.4 / NVD (CVE-2026-101356)

10.21 · Scene 11

The probes arrived six hours before the fix.

The kind of bug where waiting for a public PoC is already too late. Node.js moved the same day.

• 07:43 UTC Cloudflare observes attempts before the fix ships
• 14:00 PHP releases / 14:17 NVD publishes the CVE
• same day: Node.js 22.23.4 / 24.22.1 / 26.12.1, including the CA bundle
• anything exposed during those 24 hours needs a separate compromise check

6h17m probes before the fix

docs/00 timeline / CISA alert

10.21 · Scene 12

Patching last week does not put you on the fix.

The 12 October update covered 19.2 and later. This fix sits one release above it.

• CVE-2026-110264 / integer overflow in Onigmo / unauthenticated RCE / every release since 7.0 (2014)
• fixed: 19.5.1 / 19.4.3 / 19.3.6 / 19.2.9 / 19.1.9 / 19.0.10 / 18.11.13
• an instance updated on 12 October is affected again below 19.2.9
• no published workaround — if you cannot upgrade, remove internet reachability

10.0 CVSS · unauthenticated

GitLab patch release

10.21 · Scene 13

The distribution itself became the attack surface.

Eight of October's incidents are about how code is delivered. Provenance only says where it was built.

• LiteLLM 1.107.0: runs a stealer on any Python start, built from the lookalike Berri-AI
• [email protected]: a hostage token that wipes $HOME if you revoke it
• @subql/[email protected]: executes at install and again on import
• GhostAction: a fake security-audit.yml shipped secrets and git history out of 345 repos

345 repos swept · 16 minutes

StepSecurity (tensorlake) / StepSecurity (SubQuery) / StepSecurity (GhostAction)

10.21 · Scene 14

Environment variables were the secret store.

“Part of a series of attacks on platform companies aimed at AI inference providers” — the vendor's own words.

• customer environment variables read encrypted, some possibly decrypted
• all deployments paused; upstream providers asked to revoke keys proactively
• the job is every variable, not the one you think leaked
• SMTP passwords, webhook signing secrets and JWT_SECRET live in there too

ALL env vars to rotate

Railway disclosure

10.21 · Scene 15

One replayed packet corrupts memory.

Replaying a GSS-TSIG-signed UDP message breaks `named`. ISC states there is no workaround.

• CVE-2026-94300 / CVSS 8.1 (AV:N, AC:H) / ISC severity High / remotely exploitable
• Affected 9.11.0–9.18.50, 9.20.0–9.20.29, 9.21.0–9.21.26 (and S1) → fixed in 9.20.30 / 9.21.27
• The path is the GSS-TSIG handler: `named-checkconf -p | grep -i gssapi` tells you if it is configured
• No workaround. EOL branches (9.11–9.18) have to move to 9.20 / 9.21

8.1 CVSS · no workaround

ISC advisory / CVE.org / NVD / docs/18

10.22 · Scene 16

Six 0-days went to the public list.

ZDI sent the reports to the public LKML instead of the private security list, and Greg Kroah-Hartman replied to each one asking why.

• SMB Direct integer underflow (9.0) and a ksmbd race (9.0) — both remote code execution
• skb_zerocopy LPE, a vmwgfx out-of-bounds write, an nvmet uninitialised resource, a TCP search-domain overflow
• Mainline is unpatched: fixes sit in a v7.5 branch and -next (10-22), and no mainline commit references the ZDI ids
• Reduce exposure: stop ksmbd if unused, never expose SMB Direct to untrusted networks

6 reports sent public

Phoronix (2026-10-22) / lore.kernel.org / torvalds/linux / docs/17

10.22 · Scene 21

On an official build, it worked.

Imperva ran code on the stock Node.js 24.22.0 image from crafted Brotli data — no debugger, no instrumented build.

• Two entry paths: fetch() consuming an untrusted Brotli response, and Express 5 / body-parser 2 inflating a request body. A bare http.createServer() inflates neither
• One build was tested (24.22.0, Linux arm64); 22.x, 26.x and other architectures were not, and it took several attempts
• 24.22.1 rejects the same bytes. No in-the-wild exploitation is reported
• The NVD record carries CVSS 4.0 9.2 and no 3.1 score, so a severity-filtered CRITICAL list never shows it

24.22.0 exploited build : 24.22.1 rejects it

Imperva · Node2Shell / The Hacker News on X / Node.js advisory / docs/22 §2

10.23 · Scene 17

Same command: 125, then 104. Nothing was fixed.

The window ends 'today', so the number moves with the clock. A count measures disclosure and analysis batching, not risk.

• 125 nine hours ago, 104 now: -17 at the window edge, -4 left the CRITICAL band, +0 new
• only 26% have an NVD CPE mapping — and of the 381 CVEs published in the last 24 h, 0% do
• highest EPSS: 5.76% (php2shell 109619) — which has since dropped to 8.1, outside the band
• CNA and NVD disagree: MOVEit and TeamCity are 8.1 (AC:H) from the vendor, 9.8 (AC:L) from NVD

104 : 1 critical CVEs : known exploited

NVD API 2.0 / EPSS API / docs/13

10.23 · Scene 19

A sweep only sees what the database did.

Not by publication date, not by modification date, and not in any of twelve feeds — Infinity Core appeared in none of them.

• the CVEs were published 9/4-10/12; the PoCs landed 10/16 — a six-week gap
• their NVD records did not move in the 6-hour modification sweep; all six sit at Received
• infinity appears in 0 of the 13 feeds' titles — an RSS head is a window too; the archive is the record
• the missing axis is the disclosure channel: adding oss-sec picked up four items the other twelve missed

0 feeds that carried this

seclists / oss-sec / NVD API 2.0 (pubStartDate) / docs/21 §3

10.23 · Scene 20

Two criticals that are not CVEs.

24 hours of the GitHub Advisory Database. Both Next.js unauthenticated RCE advisories still have no CVE id.

• Next.js use cache and image-optimisation (SVG) RCE: fixed in 16.4.1 / 15.5.28
• DOMPurify 3.4.17, Pillow 13.0.0, GitPython 3.2.1, toml 5.0.1, Storybook 8.6.19
• a process that filters on CVE ids never sees these; decide on package name and patched version
• the same day named the Python CA floors: certifi 2026.10.19, pip 26.2.2

58 package advisories in 24 h

GitHub Advisory DB / certifi / pip / uv / sigstore / docs/21 §2.3

10.23 · Scene 23

A 9.9 that the list cannot show.

NVD's severity filter counts only its own (primary) CVSS 3.x score. A 9.9 attached by the CNA is a zero-row answer, even though the record carries it.

• ?cveId=CVE-2026-91637&cvssV3Severity=CRITICAL returns 0 rows; the record holds a 9.9, marked Secondary
• Same for the exploited CVE-2026-111497 (CVSS 4.0 only, 9.2) and CVE-2026-114625 (CNA 9.9)
• Ten hours: 104 → 98. Six rows out, zero in, nothing fixed
• The all-severity sweep and the vendor feeds caught both. A filtered list does not answer 'what is being exploited'

0 rows the answer for a CNA 9.9

NVD API · filtered / NVD · CVE-2026-111497 / docs/22 §1.2 / docs/13 §5.4

10.22–23 · Scene 24

Four incidents no CVE list named.

This read also collected the news and X. Four incidents were found there first, then checked against the vendor's own page. Figures are the vendors'.

• Arista VeloCloud Orchestrator CVE-2026-114625: an exploited 0-day at 9.9; any logged-in user becomes a superuser. On-prem fix 5.2.3.18 / 6.1.3.10 / 6.4.2.12
• Linux KSMBD: ZDI-CAN-33774 needs no authentication, scores 9.0, and has neither a CVE nor a mainline fix
• Capcom CAPCOM ID: 11,774,247 records. Onamae Desktop Cloud suspended up to 26,739 servers
• The press is the lead, the vendor page is the evidence. Arista’s primary page was retrieved on October 24; exploitation, authentication requirements and fixes were rechecked

4 incidents no CVE list named

Arista advisory 0199 / Capcom · 2026-10-22 / Onamae.com · 2026-10-21 / docs/22 §4

10.23 · Scene 25

No emergency patch this time.

CVE-2026-115560, published on 10-23, is a different RCE path. PHP gives no affected versions, no score and no fix, and promises a full advisory with the next routine release.

• All four conditions: set php_binary / php_serialize, switch to php at runtime with ini_set, before session_start(), with upload progress on
• Applications that never change the handler at runtime can leave the feature enabled — PHP says so itself
• The mitigation is one line: session.upload_progress.enabled=0 — the same one from 10-21, so if it is already in place, nothing more to do
• No NVD record at the October 23 check. PHP’s advisory is unchanged; the 17 recent KEV additions do not include 115560. Absence is not proof of safety

1 line the mitigation until the next release

PHP · 2026-10-23-1 / NVD · CVE-2026-115560 / GHSA-6wmf-4rv5-p3rp / docs/23

10.24 · Scene 26

Beyond the CVE list, again.

The October 24 sweep checks new and modified records, disclosure channels and domestic products. Twenty-four priority groups now have conditions, actions and sources.

• NVD, last 30 hours: 526 published / 3,531 modified. The sets overlap; do not add them
• 82 GitHub advisories in-window, six critical. Items without CVEs count too
• All 17 KEV additions since October 10, plus 14 research/official feeds and JVN
• 24 counts reviewed product/incident groups, not 24 newly confirmed breaches

24 reviewed groups

10-24 detailed snapshot / register

10.24 · Scene 27

Start with the two being exploited.

Arista VCO and Nextcloud Built-in CODE entered KEV on October 23. Both entries require forensic triage.

• VCO: a valid low-privilege account can escalate. On-prem fixes: 5.2.3.18 / 6.1.3.10 / 6.4.2.12
• Built-in CODE: update to 26.4.303 or remove it. Disabling the app is insufficient
• The vulnerable proxy.php path requires the CODE process to be stopped. Standalone Collabora is unaffected
• CISA federal due date: October 26. Preserve evidence and investigate compromise alongside patching

2 KEV additions · 10.23

Arista VeloCloud Orchestrator / Nextcloud Built-in CODE Server / 10-24 detailed snapshot / register

10.24 · Scene 28

Images, Wasm, auth, proxies.

Four boundaries, four sets of conditions. A severity label alone cannot establish whether your deployment is exposed.

• Pillow 13.0.0: FPX reaches image processing when olefile is installed, regardless of extension. The PoC uses a pinned setup
• Wasmtime -115889: 36.0.19 on the 36 line; 43.0.0+ for lines 37–42. Check untrusted Wasm compiled by Cranelift
• Laravel 12.68.0 / 13.29.0: TokenGuard bypass with particular MySQL/MariaDB configurations
• Echo 4.16.1 / 5.4.1: rejected WebSocket upgrades can bypass route-level checks, not independent upstream authorization

4 package boundaries

Pillow / Pillow / Wasmtime / Cranelift / Laravel TokenGuard / Echo Proxy middleware / 10-24 detailed snapshot / register

10.24 · Scene 29

Below the app, inside the bundle.

GitLab-named CVEs can affect Ruby users, and an OS Brotli update leaves Perl’s bundled copy unchanged.

• Ruby 4.0.9 / 3.4.12 / 3.3.13: applications compiling attacker-controlled regexp patterns are affected
• Regexp.timeout and length limits are insufficient. 4.0.8 lacks the complete fix
• IO::Compress::Brotli 0.023: affected range starts at 0.004001. Updating system Brotli alone does not fix it
• NGINX 1.30.6 / 1.31.7: five configuration-dependent issues; some execution paths also need ASLR bypass

3 runtime checks

Ruby Onigmo / IO::Compress::Brotli / NGINX Open Source / Plus / NGINX Open Source / Plus / 10-24 detailed snapshot / register

10.24 · Scene 30

Who can use the agent’s authority?

Authentication, TLS, vault isolation and model loading guard the tools and secrets an agent can reach.

• PraisonAI 4.6.59: AgentOS chat remained unauthenticated even after the older 4.5.128 fix claim
• req_llm 1.29.0: WebSocket TLS verification missing since 1.9.0. HTTP paths differ
• obsidian-tc 1.3.1: loopback still needs Origin/DNS-rebinding checks and vault isolation
• AWS Multi Model Server is EOL: no fixed release; migrate. Separately review x402 spending limits and signing origins

4 agent boundaries

PraisonAI AgentOS / ReqLLM WebSocket / obsidian-tc MCP HTTP transport / obsidian-tc MCP HTTP transport / AWS Multi Model Server / x402 SDK / x402 SDK / x402 SDK / x402 SDK / 10-24 detailed snapshot / register

10.24 · Scene 31

Include domestic business systems.

The review also covers business apps, LINE integrations and management APIs.

• Pleasanter 1.5.9.1: particular LDAP deployments on Linux/Docker and other affected configurations allow passwordless login with a valid ID
• LIFF SDK 2.31.1: DOM XSS affects 2.20.0 through 2.31.0
• JPCERT October 23 update: case E adds abuse of AI penetration-testing platforms
• Check authorization per API, token scope/expiry, rate limits and exposed admin services. IP or JA4 alone is not proof of compromise

3 domestic checks

プリザンター LDAP / LINE LIFF SDK / 国内Web/API連続不正アクセス / 10-24 detailed snapshot / register

10.24 · Scene 32

Match devices to their release line.

HPE’s official CSAF contains 72 CVEs, eight at 10.0. Industrial products and devices can also lack a fixed release.

• HPE AOS: 10.8.1.2 / 10.7.2.7 / 10.4.1.14 / 8.13.3.1 / 8.10.0.24. Match each CVE to its product line
• Mitsubishi GENESIS / ICONICS and others: fixed challenge responses, DCOM and libwebp. Follow product-specific fixes or mitigations
• Lenovo AVB: public test keys in multiple tablets; physical access required. EOL models receive no fix
• Maximum severity is not confirmed exploitation. Consider migration where a product cannot be repaired

72 HPE AOS CVEs

HPE Aruba AOS / 三菱電機 GENESIS / ICONICS等 / Lenovo Tablet AVB / 10-24 detailed snapshot / register

10.24 · Scene 33

Verify what follows the update.

Block lists, boot trust and stolen credentials need checks beyond installing a patch.

• Firefox 158.0.1 / Android 158.0.2 and other patched lines embed eMudhra blocks. Thunderbird 153.5.1 also fixes mail-auth issues
• BlackX follows administrator access. OS patching and Secure Boot DBX revocations are separate checks
• 70,000+ in the k0t report counts credentials and related data, not companies. Investigate keys and persistent access
• jxrlib JPEG XR memory bugs were reported without known CVEs or evidence of code execution. Check distributor fixes

3 trust checks

Firefox / Thunderbird / Firefox / Thunderbird / Firefox / Thunderbird / BlackX UEFI bootkit / k0t network-edge intrusion / jxrlib JPEG XR / 10-24 detailed snapshot / register

10.24 · Scene 40

HTTP2Shell. Check the bundled decoder.

The requested post is checked against primary evidence. This is the same Brotli CVE beyond Node.js, not a new HTTP/2-specific vulnerability.

• Google CNA: CVSS v4.0 9.2; same CVE as Node2Shell. Severity does not confirm real-world exploitation
• Envoy: 1.40.0 / 1.39.3 / 1.38.6 / 1.37.8, when Brotli decompression is configured
• Istio: 1.31.2 / 1.30.6 / 1.29.9; 1.29 is EOL. Check gateways and sidecars
• curl, requests, httpx: inspect the loaded decoder and automatic responses. Changing Accept-Encoding alone is insufficient

9.2 Google CNA · CVSS v4.0

Requested post / Patch the Planet · Brotli / Envoy / Istio / Verified follow-up

10.25 · Scene 34

Take the update all the way through.

Review cutoff: October 25, 16:34 JST. Retrieved afterward; newly covered information is not necessarily a new incident.

• Support open-source remediation through Patch the Planet
• Exploited exposure → urgent releases → keys and permissions → bundled copies
• Fixed-window NVD: 99 published / 5,432 modified. The sets overlap
• KEV remains October 23. No new catalog entries does not establish safety

Patch the Planet / CISA KEV / 10-25 snapshot

10.25 · Scene 35

Do not wait for the CVE list.

Fortinet urges priority hardening updates across five products. Detailed PSIRT notices are planned for November 10.

• FortiOS: 8.0.2 / 7.6.8 / 7.4.13 / 7.2.14
• Manager / Analyzer: 8.0.2 / 7.6.8 / 7.4.12 / 7.2.13
• Mail: 8.0.2 / 7.6.7 / 7.4.9. Authenticator: 8.0.4 / 6.6.11
• Match model, branch and upgrade path. Exposure controls are interim; this notice alone does not confirm exploitation

Fortinet / 10-25 snapshot

10.25 · Scene 36

Trust remains outside the browser.

Java has an October 24 emergency alert; Python ssl adds a compromised-key denylist. Routine updates alone do not prove completion.

• Java CVE-2026-114820: emergency update or official keystore/revocation mitigation; restart each runtime
• Oracle’s table lists affected versions, not the builds to install
• Python CVE-2026-115557: main merged; confirm delivery in each installed distribution
• Inventory OS stores, certifi, bundled JREs, containers and custom trust stores

Oracle / Oracle operations / Python / 10-25 snapshot

10.25 · Scene 37

Separate the DNS products.

Resolvers and authoritative servers also need review. Similar product names do not share update versions.

• BIND: 9.20.30 / 9.21.27; check GSS-TSIG replay and configuration prerequisites
• Unbound: 1.26.2 covers five DoQ, validation and load issues
• PowerDNS Recursor: 5.3.12 / 5.4.8. Authoritative: 4.9.18 / 5.0.8 / 5.1.5
• Knot DNS: 3.6.2; distinct from Knot Resolver. CVE counts do not count observed attacks

ISC / NLnet Labs / PowerDNS Recursor / PowerDNS Authoritative / Knot / 10-25 snapshot

10.25 · Scene 38

Permissions can survive the repair.

Primary ConnectWise evidence and final LiteLLM findings supersede earlier uncertainty. Provider recovery does not clear customer-side persistence.

• ConnectWise: keep connections disabled pending formal instructions; October 26 is a target
• Preserve evidence before deletion. Review service-principal sign-ins, Graph activity, new keys, owners and grants
• LiteLLM: unrevoked master key → GitHub OAuth token → CI secrets; replace early speculation
• Do not launch pip on suspect hosts first. Isolate, rebuild cleanly and revoke reachable credentials

ConnectWise / Microsoft / LiteLLM / 10-25 snapshot

10.25 · Scene 39

Check bundled copies and exceptions.

Compression, proxies, social integrations and industrial clients have different completion criteria.

• Brotli: Bun 1.4.4, brotli 1.2.1 / brotlicffi 1.2.1.0; rebuild compiled Bun executables
• Postiz: 2.25.3; registered users and integration prerequisites matter
• ATS: 9.2.16 / 10.1.5 / 10.2.1, but CVE-2026-111817 is not fixed in 9.2.16
• open62541: 1.5.7; clients importing remote OPC UA type definitions also need review

Bun / Python packages / Postiz release / Apache / open62541 / 10-25 snapshot

10.25 · 23:18 JST · Scene 41

Tailworm. Recheck the internal boundary.

A report on Tailscale and headscale: which peers may connect, and which privileges a permitted session receives.

• Tested: Tailscale 1.104.2 / headscale 0.29.5. These are not confirmed fixed versions
• The researcher’s 11 headings include configuration observations; not 11 CVEs or confirmed breaches
• Tailscale is investigating and fixing issues; no blanket SSH shutdown recommendation
• Start with enabled Tailscale SSH hosts and access rules; distinguish ordinary OpenSSH

Tailworm · researcher report / Tailscale engineering response / Tailworm special report

10.25 · 23:18 JST · Scene 42

Allowed in. More privilege than intended.

The researcher reports root-level Unix-socket forwarding and path races; the vendor is investigating post-authorization privilege handling.

• Requires SSH access and the relevant forwarding permission; not anonymous access to every node
• Privileged sockets can lead to Docker or management APIs; escalation is reported in the lab
• Limit sources and destinations, and remove unnecessary forwarding
• Check mode adds entry authentication; it does not fix escalation inside an allowed session

Tailworm · researcher report / Tailscale engineering response / Tailworm special report

10.25 · 23:18 JST · Scene 43

Headscale needs its own response.

Separate a self-hosted control plane from Tailscale SaaS. The report describes registration-path resource exhaustion and cache eviction.

• Tested: headscale 0.29.5; some registration DoS paths reportedly need no account
• The researcher reports forwarding cannot be denied there; reassess SSH grants separately from SaaS
• Review exposure, rate limits, memory limits and restart monitoring; mitigation is not a patch
• DERP client exhaustion separately requires directing a client to a hostile relay

Tailworm · researcher report / Tailscale engineering response / Tailworm special report

10.25 · 23:18 JST · Scene 44

One app entitlement. Other destinations.

The experimental conn25 path reportedly lets an entitled client select destinations through its connector.

• Requires conn25 and its client capability; not every app-connector deployment
• Inspect real destinations: connector-side LAN, cloud metadata and cluster services
• Remove unneeded entitlement and restrict destination egress, including IPv4 and IPv6
• Verify Kubernetes controls against CNI, host networking and forwarding paths

Tailworm · researcher report / Tailscale engineering response / Tailworm special report

10.25 · 23:18 JST · Scene 45

Inspect what Serve exposes.

Review Unix-socket TCP forwards, external interface binding and HTTP/2 cross-mount authorization separately.

• Find TCP forwards to privileged Unix sockets; the operator’s configuration is a prerequisite
• TS_SERVE_ALLOW_ALL_INTERFACES is a configuration note, not every Serve default
• Inspect header-based backend authorization and mounts; distinguish Funnel threat-model observations
• IP firewalls do not inspect local Unix-socket operations; remove unnecessary exposure

Tailworm · researcher report / Tailscale engineering response / Tailworm special report

10.25 · 23:18 JST · Scene 46

Inventory. Restrict. Track the fix.

Lab propagation is not evidence of a worm outbreak. Real-world Tailworm exploitation is not established by the sources reviewed.

• Inventory SSH, headscale, conn25 and Serve; reduce grants by feature prerequisites
• Confirmed fixes and dedicated CVEs/CVSS remain unresolved; 1.104.2 fixes a separate certificate issue
• If compromise is suspected, preserve evidence, isolate, revoke the device identity and rotate exposed keys
• Reimaging alone does not revoke copied identity; keep tracking official advisories

Tailworm · researcher report / Tailscale engineering response / Tailworm special report

10.25 · 23:18 JST · Scene 47

Broad review. Conditions before counts.

A fixed 16:34–23:18 JST delta complements the earlier review. Database updates are not breach counts.

• NVD: 87 published / 87 modified; the same 87 identifiers in both sets, across all severities
• 16 GitHub advisories, unreviewed; empty package metadata does not mean unaffected
• KEV: October 23, 1,756 entries; no new listing does not prove no attacks
• Trace Firecrawl, Dokploy and rotki to primary evidence; do not invent a fixed version

23:18 fixed-window collection / Nightly review

10.25 · 23:18 JST · Scene 48

A restoration target is not reconnect permission.

ConnectWise updated at 23:00 JST. Phased Cloud Backup onboarding remains targeted for October 26.

• Existing backup restoration access first, followed by re-enrollment guidance
• Keep connections disabled until individual instructions; infrastructure recovery is not full service recovery
• Stored backup assurance and investigation of live M365 access remain separate
• Continue PHP, GitLab, Brotli, exploited KEV and Fortinet work; use the role-based report

ConnectWise · Oct 25 14:00 UTC / 23:18 fixed-window collection / Nightly review

10.26 · 10:27 JST · Scene 49

10:27. What changed?

A dated record from PHP onward, plus the overnight changes. A database entry is not an incident.

• NVD: 28 published / 4,276 modified; the former are included in the latter
• 93 GHSAs, all unreviewed; many follow earlier NVD entries
• Unit 42 withdrew the PHP2Shell initial-access attribution for a government incident. AWS attempt observations are separate

10/26 collection and limits / Current report

10.26 · 10:27 JST · Scene 50

Updated. The fix was missing.

Windows .NET CVE-2026-70197: Microsoft says the October 13 updates omitted this fix. An out-of-band update supplies it.

• Published 03:50 JST, October 26; requires reuse across IPv4/IPv6 datagram sockets
• Check the matching fixed line: 8.0.33 / 9.0.22 / 10.0.14, or the documented preview
• Rebuild and redistribute self-contained apps; verify the running runtime

Microsoft official advisory / Research and exact timestamps

10.26 · 10:27 JST · Scene 51

PHP also lives inside appliances.

This morning’s JVN update includes PHP2Shell in PAN-OS on RUGGEDCOM. It does not make every Siemens product affected.

• JVN feed updated at 09:00 JST; the Siemens advisory has its own publication date
• PAN-OS: some release lines now ship fixes; other ETAs remain separate
• Starman and Plack also need checks at the HTTP processing boundary

JVN Siemens update / Siemens SSA-566512 / PAN-OS status

10.26 · 10:27 JST · Scene 52

Read by date. Return to today’s actions.

Ten independent pages, October 17–26. Observation, publication, fixes and later discoveries are labeled separately.

• Daily record (Japanese) → / 10:27 review (Japanese) →
• Continue patching public services, investigating exposure and revoking credentials
• A minimum Tailworm fix and completed ConnectWise restoration remain unconfirmed here; a target is not completion

Dated reading archive / Current actions and scope

10.26 · 14:56 JST · Scene 54

You updated. Did the running binary change?

The .NET Brotli advisory distinguishes shared-runtime, self-contained and NativeAOT deployments. The diagram is a deployment schematic.

• ASP.NET Core /HttpClient Brotli decompression is off by default; check explicit opt-in and direct API input
• Check 8.0.33 / 9.0.22 / 10.0.14. Redeploy self-contained apps; update SDK/packs and republish NativeAOT
• This supplements an advisory published at 04:06 JST. It is not a new afternoon compromise

3 deployment paths · schematic

Microsoft .NET Brotli advisory494 / 14:56 collection and conditions

10.26 · 14:56 JST · Scene 55

Released. Rolling out. Targeted.

A published patch, staged endpoint rollout and restoration target are different states. This schematic explains the distinction; verify the actual build and individual instructions.

• PAN-OS: four 11.1 hotfix states updated; CNA JSON says 14:45 JST, with a conflicting CSAF history time documented
• Neat OS26.5.4 is a staged rollout; the calendar date does not establish every device is patched
• ConnectWise’s next notice is scheduled for 23:00 JST. A target is not permission to reconnect
• 14:56 review, actions and primary sources (Japanese) →

PAN-OS official update timeline / Neat26.5.4 release notes / ConnectWise advisories

10.26 · 21:35 JST · Scene 56

Change credentials. Review the connections that remain.

GMO’s 21:00 notice extends protective checks beyond FX to securities, mail and Google accounts. This is a schematic of the checks, not live telemetry.

• Use a separate trusted device to change passwords, log out all devices and check MFA
• Review linked apps, mail forwarding, extensions and devices. Keep browser sync paused until checked
• Preventive guidance does not establish that every customer device is infected
• Official guidance and conditions (Japanese) →

GMO：不正アクセスに関する第6報（21:00 JST） / 21:35 JST 一次資料と条件

10.26 · 21:35 JST · Scene 57

A malicious release requires checking the execution environment.

New npm and PyPI malware records identify install-time or import-time execution. This response schematic is separate from a count of exploited vulnerabilities.

• Match package names and affected versions: @worca/app RCs, npm mining/persistence and the PyPI netweavex loader
• If an affected release ran, isolate the environment and revoke reachable credentials from a safe device; assess rebuilding under the advisory
• Two separate PROBABLY_PENTEST classifications do not establish the same observed harm
• Package conditions and response (Japanese) →

GitHub malware勧告：@worca/app / 21:35 JST 保存原本・対象パッケージと版

NOW · Scene 53

The risk is not the code you write.

What broke in October was the trust layer that decides which code runs. The controls that help are dull, and mostly configuration.

• release immutability, plus signature verification at install time
• pin versions with hashes; no `latest`; read the diff against the git tag
• no plaintext, long-lived admin credentials; keep the admin plane private
• assume a vendor's database holds your keys (TeamUpdraft / Railway / Zeabur)
• the kernel is the boundary you cannot patch per app: code in a container is a step to host root

docs/14 supply chain / docs/15 Apache / docs/16 snapshot 21:20 JST / docs/19 snapshot 07:25 JST / docs/20 Infinity Core / docs/21 broad sweep 10:36 JST / docs/22 wide read 20:15 JST / docs/23 PHP follow-on / Repository