> deep dives
Long-form breakdowns, incident anatomies, and opinions that don't fit in a news post.
Verification Theater: The One-Week Pattern Behind BragJack, Plugin4Shell, and WSO2's JWT Bypass
Three unrelated disclosures landed between September 13 and 19, 2026 — a browser AI agent hijack, a Git SHA-pinning bypass in every major coding agent, and a JWT auth bypass under active exploitation. All three share one root cause: a check that confirms a label matches instead of verifying the object it names.
Three Strikes: How Cisco's Firewall Brain Became Everyone's Favorite Target
In 2026, Cisco Secure FMC took three separate maximum-severity zero-days — and the third one put a Russian APT and a ransomware affiliate on the same box, in the same weeks, running the same playbook. That convergence is the story, not the CVE.
The WannaCry Bugs Never Left: Windows' Core Network Stack Just Had Its Worst Year Since EternalBlue
Netlogon, DNS Client, DHCP Server, DNS Server — four unauthenticated, network-reachable, CVSS-9.8-class memory corruption bugs in Windows' core infrastructure services in five months. This is not four unlucky patch cycles. It's a pattern, and most vulnerability-management programs are triaging it wrong.
The ECC Excuse Is Dead: A Hardening Guide for Multi-Tenant GPU Infrastructure After GPUThor
For four months, 'enable ECC' was the official fix for GPU Rowhammer. GPUThor just showed that advice was wrong. Here's what to actually change on GPU fleets that run untrusted CUDA code from more than one tenant.
Phishing the Protocol: How 2026 Attackers Made MFA Irrelevant
Device code grants, app passwords, OAuth consent screens, and WhatsApp device linking all share one property: they're real login flows, not bugs. 2026's biggest identity attacks stopped stealing passwords and started collecting the tokens MFA can't protect.
The Modem Nobody Audited: Inside the 2026 Water Utility PLC Attacks
Thirty-plus Minnesota water utilities lost control of their PLCs in a single weekend, and the entry point wasn't the internet-facing HMI everyone scans for — it was a cellular modem nobody put on the asset inventory. A look at what CyberAv3ngers actually did, why a 2021 CVE is still unpatched, and why 'get it off the internet' misses the real exposure.
Inside the AI-Orchestrated EDR Evasion Lab: What Sophos Actually Found, and Why the Bug Wasn't in the Malware
Sophos recovered a fully autonomous malware R&D pipeline — a coordinator agent, four subordinate agents, a self-provisioned lab, and 80 evasion modules built against three EDR vendors. The interesting part isn't that it worked. It's the one thing in the whole pipeline that didn't.
The Tenant Boundary Is a Fiction: Inside 2026's Cloud Cross-Tenant Bug Class
Five major cross-tenant breaks in twelve months — Cosmos DB, Vertex AI, Entra ID, AKS Backup — share one root cause: a privileged control-plane identity that trusts a customer-supplied name, key, or token it should never have accepted. Here's the pattern, and what to actually do about it.
Negative Time-to-Exploit: AI Bug Hunting Just Broke the One Assumption Your Patch Cycle Depends On
Kimi K3 found 19 Redis zero-days in 90 minutes. XBOW is #1 on HackerOne's global leaderboard. Anthropic's Mythos Preview found thousands of unpatched flaws across every major OS and browser. Meanwhile the average critical vulnerability still takes 252 days to fix. That gap is now the whole game.
Six Bulletins, One Bug Class: What Ubiquiti's 2026 UniFi Cadence Reveals About Shared-OS Edge Platforms
Since October 2025, Ubiquiti has shipped six security bulletins covering the UniFi line — five of them containing CVSS 9.9-10.0 flaws, two of them the exact same shell-injection bug class shipped seven months apart. This is what happens when one Nginx gateway and one OS layer sit in front of your network, your cameras, and your door locks.
The Agent Is the Payload: How AI Coding Agents Became 2026's Fastest RCE Pipeline
Six incidents in six weeks show the same failure mode: AI coding agents treat untrusted text as instructions and shell access as a convenience feature. Prompt injection to RCE is no longer theoretical — it's a documented, repeatable kill chain, and the guardrails vendors are shipping don't touch the actual boundary.
The Login Path Is the Target: Inside the PAM/OpenSSH Backdoor Playbook Attackers Keep Reusing
Sygnia's Operation Highland found a China-nexus group living inside an air-gapped network for a decade by backdooring pam_unix.so and sshd. It's the same target the XZ Utils and Ebury campaigns went after — because the Linux authentication stack is the softest hard target in your fleet.
Your Backup Server Is a Domain-Admin Factory: The Kill Chain Ransomware Operators Have Automated
The backup server is the highest-privilege machine in most environments and the least-hardened. Ransomware operators have known this for years and built repeatable kill chains around it. This is how they work and what it takes to stop them.
One Symlink From Host Root: The runC maskedPaths Escapes and the Myth of the Container Boundary
Three runC CVEs disclosed in November 2025 turned container escape back into a /dev/null symlink race — and one of them walks straight through AppArmor and SELinux. Here is how the maskedPaths breakout works, why seccomp and user namespaces are the layers that actually held, and what to change before the next runtime CVE.
eBPF Cuts Both Ways: The Kernel Rootkit Is Now Standard Issue in 2026's Supply-Chain Malware
In two weeks, IronWorm and the atomic-lockfile AUR compromise both shipped an eBPF kernel rootkit as just another payload module. The observability primitive your stack is built on is now the malware's stealth layer — and most detection assumptions are structurally defeated.
Anatomy of the Interlock Campaign: How a ClickFix Gang Learned to Burn Firewall Zero-Days
For a year, the surest way to get hit by Interlock was to paste a command into your own Run dialog. On January 26, 2026, the group stopped waiting for users to make mistakes and started exploiting a pre-auth, root-level Cisco firewall zero-day instead. The same crew now runs both ends of the sophistication ladder — and that should change how you model initial access.
SSRF to the Model, Model to the Cloud: The Inference Layer Is 2026's Softest Attack Surface
Model gateways and inference servers are repeating two decades of solved web-security mistakes — default-open binds, pickle RCE, pre-auth SQLi, and SSRF straight into cloud credentials. A field guide to the AI control plane's softest links and how to harden them before the next 36-hour exploitation window.
The Edge Device Audit: Turn CISA's BOD 26-02 Into a Playbook You Can Actually Run
CISA's BOD 26-02 just handed every infrastructure team a free edge-device audit checklist. Here is how to run it on your own network — inventory, version, exposure, and end-of-support triage — before an attacker runs theirs.
NTLM Coercion's Quiet Resurgence: Why 2026's Zero-Click Attacks Look Like 2021
Two unrelated bugs in the last month — an incomplete APT28 patch and an unpatched RPC defect — both hand attackers a 1990s-era credential primitive. The fact that NTLM coercion still works in 2026 is not a series of accidents. It is the model.
Worms All the Way Down: Why npm and PyPI Will Keep Spawning Self-Propagating Compromises Until We Re-Architect Install-Time Trust
From the original Shai-Hulud in September 2025 through CanisterSprawl, the Bitwarden CLI compromise, and Mini Shai-Hulud, every major npm/PyPI worm of the last eight months has used the same primitive: package lifecycle hooks that run arbitrary code on install. Until the registries change that default, each generation will keep landing.