Read Time: 9 minutes

Brought to you by:
Howdy friends!
What in the world is going on? Are you all literally drowning? So many of my friends are absolutely running on fumes with all the news hitting faster and faster. If your job includes responding to these things, itโs been near impossible to keep up.
The good(?) news is that despite all these crazy outages, most of them seem super boring and IT related, not a massive cyber attack. Though my inboxes are full of conspiracy theories there.
Please take care of your selves and your teams. The burn out is more real than ever right now.
ICYMI
๐ Something I wrote: Your pentest reporting tool has a year 2 problem*
๐ง Something I heard: A very special guest and rare pokemon sighting on The Low Down this week
๐ค Something I said: can security keep up in the age of ai?
๐ Something I read: this site that tracks newly registered 2shell domain names. hilarious but also probably, unfortunately, useful.
*Sponsored
Vulnerable News
And then nginx. F5 shipped fixes Wednesday for a pile of memory bugs, and Wiz and Xint both went public the same day. The big one, CVE-2026-123007, is a flow control math bug in HTTP/2 that's been sitting there since nginx added HTTP/2 in 2015. Unauthenticated, works with HTTP/2 on and nothing fancier than static files over HTTPS, and it leaks memory before it corrupts it, so ASLR is no help. Wiz got RCE on official builds, and Xint got RCE out of five separate nginx bugs, then chained a Linux kernel SCTP bug to get root on the host, including from inside a Docker container. (So to nit pick a bit, its not really nginx2root - its nginx + unrelated linux kernel bugs 2root)
Exploitation still takes real work (a memory leak, heap grooming, tuning per build) and nobody has seen it in the wild yet. Wiz expects attackers to go after popular packaged builds first, same as PHP2Shell. Upgrade to 1.30.7 or 1.31.9, NGINX Plus has its own fixed builds. Turning off HTTP/2 kills both HTTP/2 bugs if you need a stopgap, but it can break gRPC clients and you need a full restart, since a reload keeps existing connections alive. The kernel bug has no patch yet, so block the SCTP module anywhere you don't use it. Distros backport fixes without bumping the version number, so your distro's advisory is the source of truth. (read more here and here)

The median vulnerability takes 252 days to fix. Attackers need minutes. Legitโs remediation agents close that gap for SAST and SCA findings. Before an agent acts, each finding is enriched with reachability, ownership, and business impact, and false positives are filtered out, so only confirmed risk reaches the queue. The agents fix the root cause, open parallel pull requests across every affected service, and route each one to the right owner. Every fix is re-scanned and checked for regressions before it merges. See agentic remediation in action. (read more)
*Sponsored

Calif found a memory corruption bug in the hardware-accelerated Base64 decoder OpenJDK has shipped since 17.0.6, and the trigger is 64 A's and an equals sign. The decoder rejects that input, but only after it has already written past the end of a heap buffer. Calif turned it into Heartbleed-style memory reads and then RCE, with pre-auth code execution on default installs of Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate. Every major OpenJDK build is in scope: Oracle, Corretto, Temurin, Zulu, Red Hat, Microsoft.
The preconditions are real, though. It only fires on x86-64 chips with AVX-512 VBMI (Ice Lake and newer Xeons, AMD Zen 4 and up), ARM is clear, attacker input has to reach a Base64 decode, and every target needs its own heap grooming chain. Calif's own words: "not a universal point-and-shoot." Nothing confirmed exploited yet, but honeypots are catching scanning and blind attempts and several vendors already have working exploits. Take the November Oracle/OpenJDK updates. If you can't today, there's a JVM flag that disables the accelerated Base64 path, which is the stopgap most vendor advisories point to. (read more here, here and here)
A month after PHP2Shell, weโre back with 18 more CVEs fixed in 8.5.13, 8.4.28, 8.3.37 and 8.2.36. The scary one is a critical request smuggling bug in PHP-FPM that gets unauthenticated code execution, but only when the web server in front reuses persistent backend connections to FPM (certain nginx, Apache, HAProxy and LSAPI setups). Standard nginx and Apache setups that open a fresh connection per request are clear. There's also a handful of use-after-frees and session bugs that can reach RCE, and PHP says the critical ones all need non-default or uncommon configs.
Nothing reported exploited. Patch your branch. If you can't yet, turn off persistent connections to FPM and apply the same session upload setting PHP recommended for PHP2Shell, and keep FPM off untrusted networks, since Aikido points out the stopgaps leave some of the 18 uncovered. (read more here and here)
Vulnerabilities are found faster than ever, and we're shipping more code daily. Keeping up takes modern application security, not legacy tools that flag everything. Our agents scan your code (SAST) and dependencies (SCA), then investigate and help you fix every finding with context from your code and cloud. We even built agents for secrets, licenses, and SBOMs. That's Maze Code, and it's officially here. (read more)
*Sponsored
And Ruby makes it a full set. Gotta catch โem all! Ruby 4.0.10, 3.4.13 and 3.3.14 fix seven memory corruption bugs in the regex engine and string case conversion, and the Ruby team says it's aware of PoCs that turn several into RCE. The headliner, now CVE-2026-132825 and nicknamed ruby2รhell (yes, that's a German sharp s), is a one-byte heap overflow when Ruby changes the case of a crafted string in certain legacy encodings. UTF-8 strings don't trigger it, but Aikido and others report it's reachable without auth in default Rails and classic-style Sinatra apps. Stephan Lipp at Ada Logics and Anthropic get credit for four of the seven.
No exploitation in the wild so far, and the researchers who named it say they'll publish their exploit in about a week, so that's a deadline... The fix lives in the Ruby runtime, so bumping gems or Rails won't cover you. Check anything that bundles its own Ruby (GitLab, GitHub Enterprise Server, Mastodon, Discourse and friends) and chase those vendors, then restart or redeploy so the new runtime actually loads. Ruby 3.2 and older get nothing. Second round of Ruby regex fixes in a month, after October's GitLab-driven batch. (read more here and here)

Scott Helme and the Report URI team caught malicious JavaScript being served to visitors of 178 orgs on Monday, all through the Optimizely Web snippet that sites load straight from Optimizely's CDN. Zoom, the BBC, CNN, Salesforce, GitLab, IKEA, HSBC, Equifax, Ledger, Peacock. The payload swapped crypto wallet addresses on the page and in your clipboard, added a card skimmer for site-hosted payment forms that afternoon, and fed the Coruna and DarkSword exploit kits to older iPhones. About $7K in crypto landed before cleanup that night. Optimizely's incident report says someone abused a flaw in its account signup flow to act as admin users and write code into customer projects, and Mandiant is on it.
The file came from a domain every one of these sites had already allowlisted. Scott checked the CSPs on 256 affected hosts and only 24 blocked the payload, partly because of settings Optimizely's own docs tell customers to allow. If you run Optimizely Web, review your project code, rotate user and API credentials, treat cards typed into your own forms on Nov 16 as exposed, and rotate any session tokens JavaScript can read. If you hit one of these sites on an iPhone that day, update iOS. (read more here and here)

Microsoft's official Agent Framework package on PyPI shipped two poisoned releases Tuesday evening, 1.26.1 and 1.26.2, 33 minutes apart. The framework code itself was untouched. Each release added a startup hook that Python runs on every interpreter launch, so you didn't even have to import it, and it scooped up environment secrets, cloud and registry credential files, Azure managed identity tokens, and in the second version GitHub Actions OIDC tokens. Wiz says the attacker appears to have had credentials for Microsoft's internal signing and publishing pipeline, got bounced once for trying from AWS, then succeeded from Microsoft-owned IP space. Both versions were gone in about 55 minutes.
The attacker exfiltrated to a free-tier webhook capped at 100 requests, and Wiz found only three real victim exfils in the ~50 it could see. It's still the second Microsoft PyPI package popped this year after durabletask in May (Wiz hasn't linked the two). If 1.26.1 or 1.26.2 shows up in any lockfile, CI log or image, pin to 1.26.0 and rotate everything that environment could see, Azure DevOps PATs and anything its managed identities can reach included. Same lesson as last week's GitHub CI stealer: your pipeline's secrets are the prize. (read more)

The legend, HD Moore, and his team at runZero/Dragos spent July through November seeding about two million prompts into AI services, each carrying unique canaries (URLs, DNS names, AWS keys to empty and heavily logged accounts), then watched who touched them. Three providers turned out to be breached. In the biggest, every request during a few late-October windows got copied at the front door, and the stolen AWS keys were tested within hours, used to dig through S3 and Secrets Manager, try Bedrock, and add IAM users. That provider still hasn't disclosed. Our old friend ARTEX even showed up in CloudTrail checking whether keys were live.
Marketplace resellers just listed models at the lowest price, turned on request logging and harvested prompts. A Chinese AI lab's scraper fetched secret URLs from more than half the prompts sent to its models, including about two-thirds of the ones flagged Zero Data Retention, even when routed through US hosts. HD's point is that the privacy flag only binds the first hop, and agents will paste whatever they can read, credentials included. Seed canaries in what your own agents send, pick your upstream providers yourself, and stay out of the reseller marketplaces. (read more)
Infoblox, the DNS security company sitting in a whole lot of enterprise networks, confirmed that someone got unauthorized access to "certain company information." The way in was a compromised third-party vendor that does SharePoint data migration. Infoblox won't name the vendor, won't say how many customers are affected, and won't say whether anyone has made demands. A source told TechCrunch sensitive customer data was taken.
Infoblox says there's no sign its products, cloud services, software updates or build-and-release process were involved, which is the part that matters if you run their gear. For everybody else, it's a reminder that migration contractors get broad, temporary access that nobody remembers to take back. If you've had one through in the last year, go see what it can still reach. (read more)
Last week I joked that we took Microsoft down to give you a break. Then the FAA grounded US flights, JPMorgan payments broke, Apple Pay and iCloud hiccupped, and the NSA's classified internal network went down. Six sources told CNN NSANet was out Tuesday with no evidence of a cyberattack, and a US official later called it a "temporary degradation" during scheduled maintenance. Microsoft's was a network capacity upgrade gone wrong, the FAA's hit during a copper-to-fiber swap, and JPMorgan says software bug. Rubio's line: "We are not under systemic attack."
Bad changes are risky and its why we always code freeze during sensitive times. Guess when we usually stop all changes for a long time? The holidays. And a cluster of bad weeks is usually just a cluster of separate bad changes. What actually connects this week is patch volume. Java4Shell, nginx, PHP and Ruby all landed within days, AI-assisted research keeps surfacing bugs that have been sitting in code for a decade or more, and every one of those fixes is one more change somebody has to push to prod. Expect more weeks like this. (read more here, here and here)

Great sting here. A Florida "data recovery" company told ransomware victims it had advanced crypto tools that could restore 100% of their data without paying. Prosecutors say it was just quietly negotiating with the gangs for the decryptor and billing victims a markup. So the FBI played both sides. In June an undercover agent posed as a Rhode Island business that got hit, while another FBI employee played the ransomware crew, encrypted the fake company's files and left a note. The company charged $2,750 for a forensic assessment, quoted the "victim" $13,260 to decrypt, then emailed the "attacker" asking for the lowest possible price on the decryptor, haggled it to about $6K in bitcoin and kept the rest.
The owner, Juan Ricardo Palacio Escobar, is indicted in Rhode Island on three counts of wire fraud. The sting only netted him about $10K, but the indictment says a Kansas business paid $720K, and he sent a gang about $445K and pocketed the difference. Second ransomware middleman to land in court this year after the Florida negotiator who pleaded guilty in April to feeding BlackCat his employer's clients' negotiating positions. If a recovery vendor promises they can decrypt without paying, ask them how. (read more here and here)
Miscellaneous mattjay



How'd I do this edition?
Parting Thoughts:
Community was foundational in launching and propelling my career. Community is the only reason I can stand being in Texas during the summer months. Community is the point. Today, I invite you to embrace discomfort on the road to a more vulnerable you.
Stay safe, Matt Johansen
@mattjay


