Field Notes — Week of August 9, 2026
A short, opinionated digest of what actually mattered this week across cloud, security, AI, and automation — and what it means if you run real systems.
Security: an actively exploited N-central bypass hits the KEV
CISA added CVE-2026-18577 (CVSS 8.2) to its Known Exploited Vulnerabilities catalog on August 3 after N-able confirmed a “limited number of customers” were compromised. The flaw is an authentication bypass in N-able N-central — and, awkwardly, it’s an incomplete patch for CVE-2026-18556, the same class of bug. It’s fixed in version 2026.3 HF1, and FCEB agencies were given until August 6 to apply it.
The reason this one stings: N-central is an RMM platform, so a bypass yields admin on the server and then a pivot into every managed endpoint via the built-in Take Control feature. Huntress observed attackers reconnoitering domain controllers and moving laterally, connecting through Mullvad and NordVPN exit nodes under the default “MSP Support” username. N-able’s IOCs include a rogue svchost.exe in a user’s Documents folder and a bogus Cloudflared service used for covert tunneling.
The take: RMM is a supply chain you operate yourself — one bypass and the blast radius is your entire managed fleet. Patch to 2026.3 HF1 now, then assume compromise: hunt the published IOCs, review Take Control session logs for the “MSP Support” account since late July, and check for unexpected Cloudflared services. And note the pattern — a fix for the incomplete patch of a prior bug — as a reminder that “we patched that” is not the same as “that class is closed.”
Automation: Pulumi Cloud eats Terraform state — and HCL
Pulumi made a deliberate play for the Terraform installed base this week. Pulumi Cloud is now GA as a Terraform (and OpenTofu) state backend: teams keep using the terraform/tofu CLI for daily work but get encrypted state, state locking, update history, RBAC, audit policies, and unified resource visibility — no migration required. Its private registry can now host Terraform modules alongside Pulumi packages, usable from both Pulumi and Terraform programs. And HCL is now a first-class Pulumi language, with access to the full provider ecosystem.
This is less “rip and replace” and more “adopt the control plane without touching the config language.” It’s a direct answer to the post-BSL fragmentation, where Terraform, OpenTofu, and Pulumi users all live in the same organizations.
The take: The interesting move here is decoupling where state and governance live from which language your team writes. If you’re running Terraform or OpenTofu and feel the pain of home-rolled state buckets, DIY locking, and scattered modules, this is a low-commitment way to consolidate governance without a rewrite. Pilot it on one non-critical stack, confirm the locking and RBAC behavior matches your expectations, and treat the state backend as the Tier 0 asset it is — whoever controls state controls your infrastructure.
Cloud: Google’s Managed Airflow jumps to 3.2.2, and 2.10.5 starts its exit
A quieter week for keynotes, but a meaningful lifecycle signal: the August 5 release of Managed Service for Apache Airflow (formerly Cloud Composer) brought Airflow 3.2.2 to Gen 3 and upgraded apache-airflow-providers-google to 22.2.2. Several older builds hit end-of-support, and — the part to circle — Airflow 2.10.5 will stop being included in new Managed Airflow images and builds starting September 2026.
The take: The Airflow 2-to-3 migration has moved from optional to scheduled. If your DAGs still assume 2.x, you have roughly a month before new environments stop offering 2.10.5, and the 2→3 jump is not cosmetic — provider changes, deprecated operators, and executor differences all need testing. Stand up a Gen 3 / Airflow 3.2.2 environment now, run your DAGs against it in a lower environment, and fix the breakage on your schedule rather than on an end-of-support deadline.
AI: Kognitos wires finance rules into governed agents
Kognitos launched Context Graph for Finance on August 4, aimed squarely at the CFO’s office and starting with accounts payable. Rather than point an LLM at finance data and hope, it encodes vendors, invoices, GL accounts, cost centers, approval hierarchies, policies, and control thresholds into a single connected graph, then lets specialized agents reason and act on top of it with what the company calls deterministic, “hallucination-free” execution.
The take: This is the shape production agents are converging on — not a bigger model, but a governed substrate the agent must operate within. In finance, “mostly right” is a failed control, so the graph-plus-deterministic-execution pattern is the right instinct: encode the rules, make actions auditable, and keep a human in the approval path. Before you greenlight an AI agent for anything with money or compliance attached, ask where the rules live and how every action is logged — if the answer is “in the prompt,” it isn’t ready.
ZelMor LLC helps organizations and consultancies with cloud architecture, AI enablement, security & IAM, and automation. If any of the above is a live question for your team, start a conversation.