CloudHero Weekly Issue 001
PHP2Shell, cloud recovery, and AI agent security
When production tests the promise.
Developments from to
Sources reviewed 27 October 2026. The ConnectWise section includes a follow-up from 26 October. Availability and incident findings can change.
The 30-second briefing
What deserves your team’s attention?
- Act now if you run affected PHP: verify the runtime patch and rebuild the images your next deployment or restore will use.
- Review cloud recovery: test failover when spare capacity and management APIs are constrained.
- Review AI agent access: establish permissions, approvals and revocation outside the model before unattended production use.
A replicated service can still share a recovery dependency. A trusted backup integration can become an access path. A generally available agent can still lack the approval control your workflow needs. This cloud, AI and DevOps weekly briefing examines five developments from 19–25 October 2026 through the decisions they create for production teams.
Act now PHP security · Application and platform teams
Which PHP versions fix PHP2Shell?
PHP 8.2.35, 8.3.36, 8.4.27 and 8.5.12 fix CVE-2026-109618 and CVE-2026-109619. PHP released the updates on 21 October. Distribution packages can include backported fixes without matching these upstream version numbers, so confirm your vendor’s advisory as well as the build installed in the web-serving runtime. PHP security announcement.
AWS observed an exploitation attempt in MadPot telemetry on 17 October, before disclosure; it could not establish whether the early attempt would have succeeded against a real system. That makes the exposure window relevant to investigation, rather than proof that every vulnerable deployment was compromised. AWS threat-intelligence account.
Next step: verify the serving runtime, rebuild container images, replace deployment and recovery artifacts, and investigate suspicious activity from before disclosure. A command-line PHP version alone may not describe the PHP-FPM or Apache runtime serving requests.
One configuration caveat: PHP’s 23 October guidance separately addresses CVE-2026-115560, involving an unusual combination of session serialization changes and upload progress. Check that guidance even after applying the emergency releases. PHP session-upload advisory.
CloudHero’s take: patch completion should include evidence of what will run after the next rollout or restore. Updating today’s servers leaves that question unanswered.
Review Cloud disaster recovery · CTOs, platform leads and SREs
Multi-AZ vs multi-region: what does disaster recovery need?
Multi-zone deployment addresses some zone failures; it does not by itself demonstrate recovery from every regional or shared-service failure. Decide whether you need multi-region disaster recovery from the failure scenarios you must survive, your recovery time objective (RTO), your recovery point objective (RPO), and the operating cost you can support. AWS disaster recovery guidance.
Google’s preliminary report describes a six-hour, ten-minute London incident on 20 October. Power loss was isolated to europe-west2-b, but some workloads elsewhere in the region were affected by demand on remaining capacity and cross-zone control-plane validation dependencies. The report remains preliminary. Google Cloud incident report.
| Failure to simulate | Evidence to collect |
|---|---|
| A zone fails and replacement capacity is scarce | Can existing capacity sustain the critical service? |
| Compute creation or deployment APIs time out | Can you recover without provisioning new machines? |
| Queues accumulate during the outage | Can you replay data safely within your RTO and RPO? |
CloudHero’s take: use this incident to improve your test plan. It does not prove that another provider has the same failure path, or that every workload needs active-active multi-region architecture.
For Kubernetes teams, connect capacity planning with your node provisioning and disruption strategy; autoscaling still depends on resources being available.
Act now if affected Non-human identities · Microsoft 365 owners and MSPs
What should ConnectWise Cloud Backup customers check?
Review application credentials and permissions across the tenant, including access established through other applications. Huntress documented cases in which an abused ConnectWise application assigned itself Global Administrator and planted credentials on other apps. Disabling the original integration does not remove those additional credentials. Huntress investigation and guidance.
On 25 October, ConnectWise was targeting phased onboarding for 26 October after testing and security validation. Its 26 October update said preparations continued, without confirming that restoration access had opened. It reported no indication of compromise to existing backup archives; individual Microsoft 365 environments still required investigation. ConnectWise dated restoration updates.
Next step: follow the vendor’s current tenant-investigation guidance, keep affected connections disabled until instructed otherwise, and inspect added app credentials, role grants and sign-in evidence. Confirm your actual restore options with the provider.
CloudHero’s take: every backup, monitoring or management integration needs a named owner, a permission inventory and a tested revocation path. Include provider unavailability in backup recovery planning.
Review AI agent security · AI, data and platform teams
What permissions should a production AI agent have?
Give an agent only the identities, tools, data and network access its task requires. Enforce consequential approvals and revocation outside the model, and record tool actions as well as responses. A sandbox limits an execution environment; it does not automatically limit what an API credential can authorize.
Three announcements make these AI agent security questions timely:
- Microsoft, 19 October: expanded agent discovery and described task-scoped authority with risk-adaptive controls. Check availability for the specific control you intend to use. Microsoft agent-security announcement.
- Google Cloud, 22 October: described Agent Compute Engine microVM isolation and credentials held outside the sandbox. Customer preview is targeted for the coming months. Agent Compute Engine architecture.
- Databricks, 21 October: made Genie Code tasks in Lakeflow Jobs generally available. Release notes. Job tasks always use auto-approval, which cannot be disabled; Databricks cautions that it is not a security boundary. Job-task limitations.
Next step: before unattended execution, test an out-of-scope tool call, a prohibited write and emergency revocation. Record which external control blocks each action. Where a workflow requires human approval, verify that the platform can enforce it before execution.
CloudHero’s take: evaluate the agent’s effective authority across all its tools. Several individually narrow permissions can combine into a much broader capability.
Watch Enterprise AI · Engineering leaders and procurement
Can you use OpenAI models on Google Cloud now?
Selected enterprise customers can evaluate OpenAI models through a private preview in Google Cloud Model Garden. This is not general availability. Google’s 21 October announcement names GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, with purchases transacted through Marketplace and drawing down existing cloud commitments. Google Cloud announcement.
OpenAI’s access form describes the preview as invite-only for approved development and internal use cases, rather than production. Preview conditions. Google describes Confidential Compute and Zero Data Retention options as forthcoming.
Next step: compare access terms, region and data-handling requirements, latency, tool behaviour and cost per completed task. Confirm commitment eligibility with your commercial team. Plan a production change only against capabilities and terms actually available to you.
CloudHero’s take: procurement options are widening. Keep a workflow evaluation and regression suite between a new purchasing route and a model replacement.
Bring this to your next engineering meeting
Could we recover without our usual management APIs?
Choose one critical service. Identify the credentials, deployment APIs, spare capacity and backup provider its recovery needs. Assign an owner to test one unavailable dependency, then record actual recovery time and data loss against your objectives.
Go deeper with CloudHero
Cloud, AI & DevOps
Make time for the decisions.
Follow the weekly briefing for developments that change how you build and run production.