Apache RCE hits 64% of clouds as Solo.io ships agent substrate
8 min read · 12 sources
- CVE-2026-119618 lets unauthenticated attackers execute arbitrary code on Apache HTTP Server 2.4.55-2.4.70 via an HTTP/2 race condition, and 64% of cloud environments are exposed.
- Solo.io made Agent Substrate and kagent generally available, with OutSystems and Swissquote running the stack in production.
- Google's CodeMender agent now fixes security flaws in pull requests via slash commands, addressing the 170-day average fix time for software flaws.
- PlanetScale's open source Heroku Migrator uses trigger-based replication to move terabytes of data without logical replication or extended downtime.
- Istio 1.32.0 adds experimental AuthorizationPolicy targeting for Gateway API HTTPRoute resources and a unified east-west gateway for mixed meshes.
The day’s biggest story isn’t a new product launch or a funding round - it’s a race condition in Apache’s HTTP/2 module that leaves 64% of cloud environments running code a stranger can execute remotely. CVE-2026-119618 is unauthenticated, requires no exploit to be public yet, and the fix is a version bump most teams can ship in minutes. The rest of the newsletter is the usual KubeCon aftermath: Solo.io making its agent infrastructure generally available, Google’s CodeMender moving from flagging to fixing, and a pile of tools for people who run clusters for a living.
64% of cloud environments are running a version of Apache HTTP Server that a single unauthenticated request can turn into a remote shell.
Apache HTTP Server's HTTP/2 race condition is a patch-now situation
CVE-2026-119618 is a critical remote code execution vulnerability in mod_http2, Apache HTTP Server’s HTTP/2 module. The bug is a race condition in the assign_headers() function that leads to heap corruption, and successful exploitation lets an unauthenticated remote attacker execute arbitrary code. No exploit is public yet, but Wiz’s telemetry shows 64% of cloud environments have vulnerable resources and 51% have publicly exposed instances - that’s a lot of attack surface waiting for someone to write a PoC.
Affected versions are Apache HTTP Server 2.4.55 through 2.4.70, patched in 2.4.71 and later, plus mod_h2 2.0.8 through 2.0.42, patched in 2.0.42+. If you can’t upgrade immediately, disabling HTTP/2 support is the mitigation. For anyone running a default Apache install behind a load balancer that terminates TLS, check whether HTTP/2 is even negotiated before assuming you’re safe - the module can still be loaded and vulnerable even if you think you’re only serving HTTP/1.1.
Solo.io's Agent Substrate goes GA with production customers
At KubeCon North America 2026, Solo.io announced general availability for its enterprise Agent Substrate and the kagent developer experience layer. The pitch is infrastructure for agentic runtimes: an actor-based execution model that suspends idle agent sessions and resumes them on demand, with sandbox isolation and governed connectivity through agentgateway. If you’re running fleets of LLM-backed agents in production, this is the difference between paying for idle GPU time and only paying when an agent actually does something.
The community Agent Substrate 1.0 release follows six months of development with over 6,000 GitHub stars and 100+ contributors from 40+ organizations. Launch customers OutSystems and Swissquote are running the stack in production, which is the real signal - agent infrastructure is past the demo stage and into the “our revenue depends on it” stage.
CodeMender stops flagging and starts fixing in your pull requests
Google published a reference architecture for running CodeMender, its autonomous code-security agent, directly in Git pull requests via Cloud Run. The workflow is slash commands in the PR timeline: /verify runs an exploit simulation in a sandbox to confirm whether a finding is actually exploitable, and /fix generates a proposed code diff with a draft pull request. Findings come back with confidence scores, and verification results can change severity assessments - meaning the tool can downgrade a false positive before a human ever looks at it.
This addresses a real operational problem: the average developer fix time for software flaws has expanded from 59 days a decade ago to over 170 days in 2025. The security team flags, the developer ignores, the ticket rots. CodeMender’s model collapses that loop into a single PR comment. The architecture uses short-lived OIDC tokens rather than stored keys, so you’re not adding a new secret to rotate.
Lyft's fake payment provider tests what test cards can't
Lyft open-sourced its approach to payment testing at scale. Standard test cards from card networks can simulate a decline or an approval, but they can’t model partial authorization, latent captures, or refunds - the failure modes that actually break payment integrations in production. Lyft’s Fake Payment Provider embeds scenario logic directly into the payment token, so a simulated user can test a specific retry or fallback path safely in both staging and production.
For anyone who has debugged a payment integration at 2am because a capture arrived three hours after the authorization, this is the pattern you wish you’d built. The token carries the scenario, the provider executes it, and the surrounding retry logic gets exercised against a deterministic failure instead of a flaky sandbox.
PlanetScale migrates terabytes off Heroku without logical replication
Heroku Postgres standard tiers don’t offer logical replication or WAL access, which makes migrating large databases a multi-hour downtime event. PlanetScale built an open source Heroku Migrator that sidesteps the limitation entirely. The tool uses trigger-based replication to record changes on every table and replay them on the target, providing a replication-based experience without extended downtime. Version 2.0 splits the process into data transfer and index rebuild phases, which increases copy speeds by roughly five times.
For teams stuck on Heroku’s standard plans, this is the difference between a scheduled maintenance window and a live migration. The tool is open source, so you can run it against any target, not just PlanetScale.
OpenBot: a self-hostable workspace for AI coworkers
CopilotKit’s OpenBot is a modular, MIT-licensed template for deploying AI agents as team members. Each agent gets its own browser, files, and tools, and you can choose models from OpenAI, Claude, Gemini, or open models via OpenAI-compatible endpoints. Existing agents can be brought in through AG-UI, and tools connect via MCP. It ships as a self-contained desktop app for Mac and Windows, plus Kubernetes deployment options for teams that want it in their own cluster.
OpenBot for Business adds company sign-in, tool permissions, approvals, and audit - the governance layer you need before letting an agent touch production. It’s built on CopilotKit, which the team says is used in production by a majority of Fortune 500 companies, so the underlying pieces have some mileage.
TensorFold: a Zig-based LLM inference engine
TensorFold is an LLM serving engine written in Zig that runs on Apple Silicon and NVIDIA GPUs. It reads checkpoints and runs Metal or CUDA kernels directly, with no Python or MLX installation required. Install via Homebrew with brew install ashhart/tensorfold/tensorfold, and the recommended starting point is Qwen3.5-2B, a small 4-bit checkpoint tested on Metal and GB10.
The interesting bit is the Agent Traffic Control (ATC) system built in for coordinating coding agents working on the project itself - commands for checking, claiming, and pulsing issues. It’s a dogfooding story for AI-assisted development, and a Zig binary that serves LLMs without a Python runtime is genuinely lean for edge deployments.
Open Cluster Management graduates to CNCF incubating
The CNCF Technical Oversight Committee accepted Open Cluster Management as an incubating project, moving it up from the Sandbox it joined in November 2021. OCM uses a hub-and-spoke architecture with open APIs to manage Kubernetes clusters across cloud, on-premises, and edge. The asynchronous pull model maintains cluster autonomy without centralizing credentials, which matters for edge deployments where connectivity is intermittent.
Originally developed by IBM and Red Hat, it hit 1.0 in June 2025 and has since delivered updates through 1.4 in October 2026. Key features include CEL-based cluster selection, an Argo CD pull model add-on, and a pluggable add-on framework. eBay reportedly uses it with Argo CD across more than 190 clusters - a solid real-world reference for multi-cluster management at scale.
Linkerd's independent threat model found 30 threats, one bug
ControlPlane, a cloud native security consultancy, built an independent threat model of open source Linkerd and found 30 threats, 28 of them medium or high severity, plus 33 recommendations. The striking part: only one finding is an actual software bug - CVE-2025-43915, a metrics issue fixed in edge-25.2.1 by turning off affected metric labels by default. Everything else stems from default settings, operational processes, underlying Kubernetes cluster configuration, and surrounding tools like ingress controllers.
That’s a useful data point for anyone evaluating service meshes: the code is solid, but the defaults and the surrounding ecosystem are where the risk lives. The review maps threats to MITRE ATT&CK and NIST 800-53 frameworks, though it didn’t include penetration testing or code audits.
GitHub secret scanning now watches Actions workflow logs
GitHub’s secret scanning now detects secrets in Actions workflow logs, available in public preview for enterprises on GitHub Enterprise Cloud with Secret Protection or GitHub Advanced Security. When a workflow run completes, GitHub scans the logs for provider patterns and raises repository alerts with links back to the offending log. The feature is disabled by default and must be enabled by enterprise owners in Advanced Security policies.
The caveats matter: only workflow runs completed after enabling the feature are scanned, and generic patterns, custom patterns, AI-detected secrets, and already-masked secrets aren’t detected. If a secret leaked in a log from last month, this won’t find it retroactively.
Octopus Deploy kills its AI Assistant
Octopus Deploy deprecated its AI Assistant, the chat-based interface in the web portal that offered best practice suggestions, self-support for failed deployments, and project creation from prompts. The Chrome extension is no longer available to new users, and existing users see a static deprecation notice. The recommended migration path is the MCP interface, particularly the remote MCP interface, with example prompts for creating Kubernetes projects and building dashboards.
The Claude Agent Step can also automate the Claude Code CLI and interact with the Octopus MCP interface for AI features in deployment processes. It’s a quiet admission that a bespoke chat UI was the wrong shape - MCP is the integration point now.
Istio 1.32.0 brings Gateway API authorization and east-west unification
Istio 1.32.0 is out, managed by release managers from Microsoft, Tetrate, and Solo.io. The headline feature is experimental AuthorizationPolicy targeting Gateway API HTTPRoute resources via targetRefs, which finally lets you attach authorization policies to routes rather than just gateways and services. TLS tuning on gateways gets alpnProtocols and ecdhCurves configuration, and there’s a unified east-west gateway for mixed ambient and sidecar meshes.
Traffic management additions include utilization-based load balancing using ORCA load reports, richer outlier detection with configurable HTTP error codes and failure percentage thresholds, and configurable inbound retry policies. Security features include the ability to mark JWT rules as REQUIRED in RequestAuthentication. For anyone running mixed ambient and sidecar meshes, the unified east-west gateway removes a real operational headache.
You May Also Like
Gheist’s GitHub token heist stalls, Cloudflare maps Q3 outages, and autoscalers finally clash
Wiz’s research exposed “Gheist,” an automated campaign that slipped secret‑exfiltrating GitHub Actions into 850+ accounts before GitHub’s new token controls …
AWS Continuum GA, Claude on Google Cloud, and a critical Brotli RCE
AWS rolled out Continuum for code vulnerabilities, automating exploit verification and PR generation, while Google Cloud added an Anthropic‑managed Claude tier …
Spacelift Polices Cursor Origin, S3 Hits Vercel, and Cloudflare Drops Edge Log Transformers
Spacelift now treats Cursor Origin like any Git forge, applying signed Open Policy Agent guardrails to AI-generated pull requests before raw infrastructure …




