<p><strong>TLDR:</strong> The GitHub breach attributed to Team PCP is worse than initially reported: 5,718 malicious commits hit 5,561 repos via compromised PATs, and a poisoned VS Code extension exfiltrated ~38,000 internal GitHub repos. IBM and Red Hat responded to the broader supply chain threat with Project Lightwell, a $5B initiative deploying 20,000+ engineers and AI tooling to backport security patches directly into enterprise software supply chains. AWS rebuilt 97% of OpenSearch Serverless from scratch to handle bursty AI agent workloads, with scale-to-zero and 60% cost savings as the headline claims. Replay 2026 continued producing dense Temporal content, with production case studies from Twilio, Netflix, LinkedIn, and OpenAI.</p>
<hr>
<h2>GitHub breach expands: 38,000 internal repos exfiltrated, 5,718 malicious commits across public repos</h2>
<p>The GitHub security situation is significantly worse than yesterday’s summary. Team PCP’s Megalodon campaign pushed 5,718 malicious commits across 5,561 repositories using compromised PATs, stealing CI secrets, cloud credentials, SSH keys, and OIDC tokens. Separately, a GitHub employee downloaded a poisoned NX Console VS Code extension (compromised via the Tanstack supply chain), leading to exfiltration of roughly 38,000 private internal repos and customer support data. Team PCP is selling the stolen GitHub source code for $50,000 minimum. The practical risk even without customer repo access: internal code reveals architecture, secrets handling patterns, and security controls that enable targeted follow-on attacks. Audit your VS Code extensions, rotate CI/CD credentials, and treat your own source code as potentially exposed. <a href="https://app.daily.dev/feed-by-ids?id=QXAvNTbqv&amp;id=MNJTIH6ZL&amp;id=14FOK3Vdf" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>IBM and Red Hat launch Project Lightwell, a $5B open source security clearinghouse</h2>
<p>IBM and Red Hat announced Project Lightwell, a $5B initiative positioning itself as a centralized clearinghouse between vulnerability discovery and remediation for enterprise open source. The core mechanic: 20,000+ engineers plus AI-assisted triage backport validated patches directly into enterprise supply chains from configuration manifests like pom.xml, without requiring source code access or forcing major version upgrades. It launches with Java/Maven support, expanding to PyPI, npm, and Go. Early adopters include Bank of America, Goldman Sachs, JPMorgan Chase, and Visa, with commercial subscription expected within 30 days. The stated motivation is that frontier AI models are compressing attacker timelines — Anthropic’s Project Glasswing reportedly found 23,000+ security flaws across 1,000+ open source projects — and the current patch cycle isn’t keeping up. <a href="https://app.daily.dev/posts/HD6PN5yrW" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>AWS rebuilt 97% of OpenSearch Serverless for AI agent workloads</h2>
<p>AWS rewrote nearly all of Amazon OpenSearch Serverless, separating storage and compute onto a new proprietary storage layer. The driver was straightforward: AI agent workloads are bursty — idle for long stretches, then hammering the system — and the original architecture couldn’t handle that gracefully. The claimed numbers are provisioning in seconds (roughly 20x faster), true scale-to-zero, and up to 60% cost savings versus peak-provisioned clusters. Whether those savings hold in practice depends heavily on actual usage patterns, but scale-to-zero alone addresses a real pain point for teams running agents that aren’t active around the clock. The new generation supports both full-text and vector search, with native integrations for Vercel and Kiro IDE. Generally available now across all AWS commercial regions. <a href="https://app.daily.dev/feed-by-ids?id=ZpIy5U0cO&amp;id=KmxUThjV0" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>Replay 2026: Twilio, Netflix, LinkedIn, and OpenAI production Temporal case studies</h2>
<p>The Replay 2026 conference continued producing dense production content. Twilio detailed building a broadcast messaging system hitting 20,000 messages/second with fewer than 50 Temporal actions, using progressive batching to reduce action cost from O(N) to O(processing time). Netflix shared a six-year Temporal adoption journey now serving 150 teams across 500+ namespaces, running active-active across four regions. LinkedIn described migrating 4+ million CPU cores across 2,300+ stateless services to Kubernetes using Temporal for orchestration, with a supervisor hierarchy managing concurrent migrations at scale. OpenAI’s infrastructure engineer detailed 60x growth in one year to 700+ namespaces and 100+ workers, reducing time-to-first-workflow from 1-2 weeks to under a day — all managed by a team of four engineers. <a href="https://app.daily.dev/feed-by-ids?id=X7f8EZTUi&amp;id=AbyqvqOoE&amp;id=eOQJeotxK&amp;id=1JfOu9L07" target="_blank" rel="noopener nofollow">Read more</a></p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>GlassWorm botnet takedown follow-up:</strong> CrowdStrike, Google, and Shadowserver severed all four C2 channels of the GlassWorm botnet simultaneously on May 26, disrupting a two-year supply chain campaign that poisoned 300+ GitHub repos via malicious VSCode extensions and npm/Python packages. Security experts caution this is a disruption, not eradication — attackers can reconstitute quickly. Audit any packages or extensions installed during the window GlassWorm was active. <a href="https://app.daily.dev/posts/KPiHKcOdV" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>GitLab patches high-severity Duo AI identity spoofing flaw:</strong> GitLab released 19.0.1, 18.11.4, and 18.10.7 addressing 7 CVEs including a CVSS 8.2 improper access control flaw in Duo AI workflow runners that could allow identity spoofing. Self-managed installations should upgrade immediately. <a href="https://app.daily.dev/posts/Zr6CDSlKy" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>OpenTofu 1.12.0: dynamic prevent_destroy and destroy=false:</strong> OpenTofu 1.12.0 ships dynamic prevent_destroy driven by input variables (e.g., var.environment == “production”), a destroy=false lifecycle option for removing resources from state without destroying infrastructure, parallelized provider downloads during init, and a -json-into flag for structured output alongside normal terminal output. Terraform never shipped the dynamic prevent_destroy despite long-standing requests. <a href="https://app.daily.dev/posts/5Xfnrk40J" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Elastic Stack security patches across three release lines:</strong> Elastic Stack released 8.19.16, 9.4.2, and 9.3.5 simultaneously, all primarily addressing potential security vulnerabilities. Users on any 8.19.x, 9.4.x, or 9.3.x should upgrade to the latest patch in their line. <a href="https://app.daily.dev/feed-by-ids?id=1XIFSjSCu&amp;id=8MhRQHjEP&amp;id=awF9KSeS0" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Microsoft Azure Linux 4.0: first general-purpose server Linux distro:</strong> Microsoft announced Azure Linux 4.0, a Fedora-based general-purpose server distribution for Azure VMs — the first time Microsoft has offered a supported Linux distro beyond container hosting. Azure Container Linux (Flatcar-based, immutable) is also now GA. Microsoft is contributing upstream to Fedora rather than forking. Two-year support lifecycle, WSL support planned. <a href="https://app.daily.dev/feed-by-ids?id=D5hUCqswg&amp;id=7bCcwWhHs" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Canonical Workshop: sandboxed dev environments with LXD:</strong> Canonical released Workshop, an open-source tool that spins up sandboxed development environments in unprivileged LXD containers from a single YAML config file. Host resource access requires explicit grants. SDKs (Ollama, NVIDIA CUDA, AMD ROCm, OpenCode) are versioned and composable. The strict isolation model is positioned as useful for AI agent workflows where you want agents confined to containers that can’t touch the host without explicit permission. <a href="https://app.daily.dev/posts/ml3hJXlj8" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Cloudflare adds Claude Managed Agents support:</strong> Cloudflare now supports Claude Managed Agents, decoupling Anthropic’s orchestration from the execution layer. Enterprises get sandbox controls, VM-based execution, private service access via Cloudflare Mesh and Workers VPC, browser session monitoring, and credential injection. An open-source deployment template is available on GitHub. <a href="https://app.daily.dev/posts/E3p1P6YWd" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Spark 4 migration: ANSI SQL mode on by default, Scala 2.12 dropped:</strong> Apache Spark 4 drops Scala 2.12 support, requires Java 17+, enables ANSI SQL mode by default (turning previously silent NULLs into runtime exceptions), and switches structured streaming state store to RocksDB. New capabilities include a VARIANT type, SQL pipe syntax with |&gt;, and procedural SQL scripting. Recommended migration path: enable ANSI SQL mode in Spark 3 first to surface hidden errors before upgrading. <a href="https://app.daily.dev/posts/0pY2476dF" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>MoE inference economics: only favorable at high GPU utilization:</strong> A detailed breakdown of Mixture of Experts inference costs finds MoE only beats dense models when GPU utilization is consistently high — at saturation, MoE can cut per-token costs by ~50%, but at low utilization the larger memory footprint makes it more expensive. The recommendation: start with a dense model, measure real utilization for a month, then decide. <a href="https://app.daily.dev/posts/e4tW8AiL5" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Hedged requests cut p99 latency 74% in fan-out architectures:</strong> With 100 downstream services each at 1% straggler rate, 63% of top-level requests get delayed. Adaptive hedged requests — fire a backup when the primary exceeds current p90, use whichever responds first, cancel the other — drop p99 from 65ms to 17.3ms with 8.9% overhead. A token-bucket budget caps hedge rate at a configurable percentage to prevent load amplification during genuine outages. Reference Go implementation available as an open-source HTTP RoundTripper and gRPC interceptor. <a href="https://app.daily.dev/posts/FzQj0VvPB" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Hetzner third price increase in three months, effective June 15:</strong> Hetzner’s third pricing change in three months hits new orders and rescales of dedicated servers and cloud plans on June 15, 2026. The advice: build an exposure report classifying workloads by resize frequency and migration risk before making any moves. Staying on Hetzner may still be correct for stable fleets with existing contracts. <a href="https://app.daily.dev/posts/AIevBnXG3" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>YC dev tools repo security scan: median score 21/100:</strong> A scan of 6,195 public GitHub repos across 128 YC-backed developer tools companies found median security score of 21/100, no company exceeding 80. Only 2 companies (Apollo and Formance) both enable branch protection and require checks before merging. Most either have no merge gate at all, or have branch protection without enforcing any status checks. <a href="https://app.daily.dev/posts/aiVxwEX7a" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>


Server-side development, system design, API patterns, and distributed architecture.

Backend Digest

A security-heavy digest covering multiple major incidents and initiatives. The GitHub breach via Team PCP's Megalodon campaign is worse than reported: 5,718 malicious commits across 5,561 repos, plus ~38,000 internal GitHub repos exfiltrated via a poisoned VS Code extension. IBM and Red Hat launched Project Lightwell, a $5B initiative using 20,000+ engineers and AI tooling to backport security patches into enterprise supply chains, targeting Java/Maven first with PyPI, npm, and Go to follow. AWS rebuilt 97% of OpenSearch Serverless for bursty AI agent workloads, claiming 60% cost savings and scale-to-zero. Replay 2026 produced production Temporal case studies from Twilio, Netflix, LinkedIn, and OpenAI. Also notable: GlassWorm botnet disruption, GitLab Duo AI identity spoofing CVE, OpenTofu 1.12.0 with dynamic prevent_destroy, Apache Spark 4 migration notes, and a YC dev tools security scan finding a median score of 21/100.

GitHub breach widens, IBM launches $5B open source security clearinghouse