Saturday, September 26, 2026|Issue 8|6 min read|Join the discussion
Sometimes the strongest security control is simply turn it off
Kiteworks issues an extraordinary precautionary shutdown warning while Microsoft pushes Copilot toward persistent autonomous work, forcing enterprises to think harder about authority, identity and operational control.
The short version
- Kiteworks told customers to shut down their systems this weekend after receiving credible threat intelligence from federal authorities. The company says it has no indication that Kiteworks or customer systems have been compromised, and no confirmed zero-day has been named.
- Microsoft rebuilt Copilot around three destinations: Home, Code and Autopilot. Autopilot is a persistent, proactive agent that keeps working when you are not, expanding to private preview at the end of the month.
- The federal deadline for two actively exploited flaws, WSO2's CVE-2026-5430 and Adobe Commerce's CVE-2026-71362, lands tomorrow, September 27.
- Google Cloud's Security Command Center renamed AI Protection finding categories this week, a small edit that signals how seriously cloud providers now treat AI assets as first-class inventory.
- ISC2's AI security certification is still in development, with its job task analysis survey open through September 28. It is not available yet.
Issue at a glance
1 critical2 high1 notable
- criticalSecurityKiteworks tells customers to turn it off
- highAIMicrosoft's new Copilot keeps working after you leave
- highVulnerabilityTwo KEV deadlines land tomorrow
- notableAIGoogle Cloud keeps folding AI into the security inventory
Patch clock
| CVE | Product | What it is | Severity |
|---|---|---|---|
| CVE-2026-5430 | WSO2 API Manager, API Control Plane, Traffic Manager, Universal Gateway | Path traversal and JWT authentication bypass allowing unrestricted file upload and remote code execution; actively exploited | Critical |
| CVE-2026-71362 | Adobe Commerce and Magento | Incorrect authorization vulnerability; actively exploited | Critical |
Also on the vulnerability desk
- Not a CVE, but the weekend's biggest operational ask: Kiteworks recommends a precautionary shutdown window for customer systems after receiving credible threat intelligence from federal authorities. All known vulnerabilities are addressed in current release 9.5.1, and the company says it has no indication of compromise.
- Also added to KEV this week: Microsoft SharePoint code injection (CVE-2026-65660) and MikroTik RouterOS pre-authentication SSH bypass (CVE-2026-67279), plus a WordPress Core remote file inclusion flaw (CVE-2026-87902) due September 28.
Breaking
vendor advisory / threat intelligence / file transfer
Kiteworks tells customers to turn it off
Kiteworks, the secure file-transfer company formerly known as Accellion, emailed customers Friday urging a precautionary shutdown window this weekend after receiving what it calls credible threat intelligence from federal intelligence authorities that a threat actor may attempt to target some Kiteworks systems. Self-managed customers, on-premises or on AWS and Azure, were told to shut systems down themselves; Kiteworks is shutting down the systems it hosts on customers' behalf.
Why it matters. This is an extraordinary step for a vendor whose entire product is moving sensitive files, and whose customer list skews toward governments, financial institutions and large enterprises. Read the language carefully: this is credible threat intelligence, not a confirmed compromise and not a confirmed zero-day. Kiteworks says all known vulnerabilities are addressed in release 9.5.1 and that it has no indication its or customers' systems have been breached. The memory of the 2021 Accellion FTA breach, which cascaded through dozens of organizations, is doing a lot of work in how seriously customers will take this.
What to do. If you run Kiteworks, follow the vendor's shutdown guidance for your time zone and confirm you are on release 9.5.1. Treat the window as an opportunity, not just an inconvenience: review logs for unusual transfer activity before and after, and make sure you can reach the vendor's technical support if follow-on guidance arrives. If you do not run Kiteworks, note the precedent: a vendor asking customers to power down on threat intelligence alone is a new bar for precautionary response.
AI / Microsoft / agents / governance
Microsoft's new Copilot keeps working after you leave
Microsoft unveiled a redesigned Copilot built around three destinations: Home, which merges Chat and a delegated-work feature called Cowork; Code, which lets anyone build apps and automations with GitHub Copilot technology underneath; and Autopilot, a persistent, proactive agent that keeps working even when you are not. Autopilot, a rebrand of the Scout personal agent announced at Build, expands to private preview at the end of the month. Microsoft also announced FinOps for AI capabilities to manage spend across Copilot and agents.
Why it matters. A persistent agent is not a chatbot with better marketing. It has an identity, a memory, a computing environment and a workspace inside your tenant, and it acts while unattended. That makes it a machine identity in the most literal sense, and it inherits the hardest questions in identity management: what is it allowed to do, who approved that, and how do you take it away. Microsoft CEO Satya Nadella framed it himself this week: every agent has to have an identity, and everything it does needs to be observed.
What to do. Before Autopilot reaches your tenant, decide how agents get provisioned, reviewed and deprovisioned, the same joiner-mover-leaver discipline you apply to people and service accounts. Map what data and applications a persistent agent can reach through Copilot's grounding in Work IQ and the new plugin registry. And watch the FinOps controls: an agent that works all night also bills all night.
KEV / patch deadline / WSO2 / Adobe
Two KEV deadlines land tomorrow
Federal agencies have until Sunday, September 27, to remediate two actively exploited vulnerabilities CISA added to the KEV catalog on September 24: CVE-2026-5430, a maximum-severity flaw in WSO2 API Manager, API Control Plane, Traffic Manager and Universal Gateway, and CVE-2026-71362, an incorrect authorization flaw in Adobe Commerce and Magento. The WSO2 flaw stems from the JWT authentication mechanism accepting tokens signed with an unsupported algorithm; watchTowr says its honeypots captured exploitation attempts using forged JWT tokens as early as September 13.
Why it matters. WSO2's products sit in front of APIs and identity flows, which means a bypass here is a bypass of the thing that decides who gets in. Adobe Commerce and Magento sit in front of payment flows. Both are authority concentrators, and both are now on a clock. There has been some disagreement among researchers about how to classify the WSO2 flaw, path traversal versus authentication bypass, but the practical answer is the same either way: patch it.
What to do. If you run affected WSO2 versions (API Manager 4.1.0 through 4.6.0, Traffic Manager and Universal Gateway 4.5.0 and 4.6.0) or Adobe Commerce and Magento, apply the vendor fixes today rather than tomorrow. Per BOD 26-04, federal agencies should also check whether the system was compromised before the patch went on. Everyone else should treat the KEV deadline as the free prioritization signal it is.
CVE-2026-5430, CVE-2026-71362
cloud / AI security / Google Cloud
Google Cloud keeps folding AI into the security inventory
Google Cloud's Security Command Center renamed a set of AI Protection finding categories this week, dropping the legacy VERTEX_1P_ prefix and clarifying that AI Protection detects tuned models. It follows a September 16 preview of MCP server endpoints that let LLM agents perform investigative and management tasks inside Security Command Center itself.
Why it matters. Individually these are small release-note items. Together they show the direction: models, agents, prompts and inference endpoints are being treated as first-class cloud assets with their own inventory, findings and attack-path simulation, not as an application detail buried inside a VM. If your cloud security posture management does not know what AI assets you have, it does not know what you have.
What to do. If you are on Google Cloud, check whether AI Protection is activated at the organization level and whether your AI asset inventory matches what your teams think they have deployed. If you are on another cloud, ask the same question of your own tooling.
Trends
Availability is still a security control
Kiteworks did not patch, mitigate or workaround. It asked customers to turn the system off. That is worth sitting with. When threat intelligence outpaces disclosure, when a vendor knows something is coming but cannot yet say what, reducing exposure to zero for a few hours is a rational, arguably elegant control. The industry reflex is to keep everything up at all costs, but availability is a business requirement, not a security virtue. A system that is off cannot be exploited, and a team that can gracefully turn things off has more options than a team that cannot. The uncomfortable question this weekend raises for everyone else: if your vendor called you Friday afternoon with the same request, could you do it?
Persistent agents are becoming machine employees
Microsoft's Autopilot is the clearest signal yet that the big platforms intend agents to be persistent actors inside the enterprise, not session-based helpers. A persistent agent holds objectives, accumulates memory and acts while unattended, which makes it look less like software and more like a staff member with a badge. That framing is useful, because enterprises already know how to govern staff: joiner-mover-leaver. Provision deliberately, review access when the role changes, deprovision completely at the end. The organizations that struggle with agents will be the ones that treat them as features to enable rather than identities to govern.
AI FinOps is becoming part of governance
Microsoft paired its agent announcements with new FinOps for AI capabilities to manage spend across Copilot and agents, and that pairing is not accidental. An agent that works autonomously consumes compute autonomously, which means cost, identity and authorization are now the same conversation. Governance for persistent agents is not just what can it access, it is what can it spin up, what can it spend and who gets the bill. Expect AI cost controls to show up in security reviews sooner than most teams planned.
Vulnerability intelligence is becoming operational intelligence
This week's KEV additions came with the now-familiar BOD 26-04 machinery: deadlines, forensic triage requirements, exposure evaluation. Meanwhile a vendor shutdown advisory arrived with no CVE attached at all, just credible intelligence and a recommendation to power down. The common thread is that a CVE record alone is close to worthless as an input. What matters is the composite: the CVE, the KEV flag, the vendor advisory, the threat intelligence, your exposure, the identity the asset holds and the telemetry you would check first. Knowing a vulnerability exists is getting cheap. Knowing whether it matters in your environment, and what to do before the details are public, is the valuable part.
Clearance corner
The federal deadline machine keeps running
BOD 26-04's risk-driven remediation model is now the backdrop for every KEV cycle: agencies prioritize flaws on publicly exposed assets that grant total control, check for compromise before patching, and defer lower-risk work. This week's two Sunday deadlines, WSO2 and Adobe Commerce, are exactly the profile the directive was written for. For cleared and federal-adjacent teams, the practical takeaway is that the triage-and-verify rhythm is now permanent, and the Kiteworks advisory is a reminder that federal threat intelligence sometimes reaches vendors before it reaches any public channel.
ISC2's AI security credential is taking shape, slowly
ISC2's AI security certification remains in development, not available, but the machinery is moving: more than 7,500 volunteers from over 145 countries have signed up, job task analysis workshops ran in Alexandria, London and Singapore in August, and the validation survey stays open through September 28. For workforce planners, the signal is that a vendor-neutral AI security benchmark is coming, and the people who help define it now will shape what hiring managers ask for later.
What we are watching next
Kiteworks follow-on
Whether the weekend passes quietly, and whether any CVE, IOCs or post-incident detail emerges from the federal intelligence that triggered the shutdown advisory.
Sunday's KEV deadline
WSO2 CVE-2026-5430 and Adobe Commerce CVE-2026-71362 remediation is due September 27 for federal agencies; watch for exploitation detail after the deadline passes.
Autopilot governance
Microsoft's private preview expands at the end of the month; the details that matter are identity, observability and kill-switch controls, not the demos.
AI asset security
Cloud providers are folding models, agents and prompts into security posture management; watch whether AI Protection-style inventory becomes table stakes across CNAPP.
ISC2 credential development
The job task analysis survey closes September 28 and item-writing workshops run through October; an exam outline should follow.
Sources
Join the discussion
Loading comments…