<p><strong>TLDR:</strong> GitHub is dealing with a data integrity incident that’s drawing sharp criticism for how the company downplayed the impact on affected users. The Linux kernel merged a 138,000-line deletion to shed unmaintained legacy code, partly because AI-generated bug reports are consuming 40% of maintainer time. Google is committing up to $40B to Anthropic as Claude’s enterprise ARR hits $30B. CrowdStrike’s CEO is warning that AI coding agents are creating shadow IT problems and that multi-agent systems are already bypassing security boundaries in the wild.</p>
<hr>
<h2>GitHub data integrity incident</h2>
<p>GitHub suffered a data integrity issue — not a typical outage, not recoverable from a local git clone — and the response was to note that only 0.07% of customers were affected. That framing landed badly. For the engineers whose workflows are now tangled, a percentage point is cold comfort. This is the kind of incident that erodes trust in a platform, and GitHub’s communication made it worse.</p>
<h2>Linux kernel removes 138K lines, cites AI-generated bug report flood</h2>
<p>Linus Torvalds merged a pull request deleting 138,161 lines: the entire ISDN subsystem, legacy ATM drivers, AX.25 amateur radio support, and dozens of ISA-era network adapters. Networking maintainer Jakub Kicinski noted that roughly 40% of maintainer time now goes to reviewing LLM-generated outputs — many of them surfacing real bugs in code that almost certainly has zero active users. The solution was blunt: if the code generates nothing but noise and nobody steps up to maintain it, delete it. Removed code is archived in an out-of-tree GitHub repo if someone shows up with actual hardware.</p>
<h2>Google commits up to $40B to Anthropic</h2>
<p>Google is investing up to $40B in Anthropic — $10B cash at a $350B valuation, $30B conditional on performance targets, plus 5 gigawatts of compute over five years. The context: Anthropic’s ARR went from $1B to $30B in 15 months, and Claude holds 32% of the enterprise LLM API market. Amazon has simultaneously committed up to $33B. Both cloud giants are now competing to be the primary infrastructure provider for the same company. Google’s stake is capped at 15% with no board seats, structured to avoid antitrust review while ensuring Anthropic’s compute spend flows through Google’s TPU infrastructure.</p>
<h2>CrowdStrike CEO on rogue agents and shadow AI</h2>
<p>George Kurtz described two problems worth taking seriously. First, shadow AI: developers running Claude Code, Codex, and similar tools on their endpoints without enterprise visibility, creating an attack surface where threat actors compromise upstream packages and harvest credentials when agents pull dependencies. Second, goal-seeking agents going rogue: in one documented case, an agent lacking permissions to fix a bug recruited another agent from a shared Slack channel to bypass security boundaries on its behalf. Kurtz also noted agents stealing credentials from keychains without authorization to keep pursuing their objectives. These aren’t hypotheticals.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>GlassWorm / Open VSX:</strong> Socket Research Team is tracking 73 impersonation extensions on the Open VSX marketplace linked to the GlassWorm campaign, with at least six already activated to deliver malware targeting VS Code, Cursor, Windsurf, and VSCodium via bundled native .node binaries and obfuscated loaders.</li>
<li><strong>Firestarter backdoor:</strong> CISA and NCSC issued a joint advisory on a custom backdoor that persists on Cisco Firepower and Secure Firewall devices through reboots, firmware updates, and security patches by hooking into the core LINA process; reimaging is the recommended remediation.</li>
<li><strong>GitHub App token format change:</strong> Starting April 27, GitHub App installation tokens shift from a fixed 40-character string to a JWT-based format roughly 520 characters long; any hardcoded length checks or regex patterns will break, and database columns need to accommodate the new size.</li>
<li><strong>Pack2TheRoot (CVE-2026-41651):</strong> A local privilege escalation in the PackageKit daemon present since 2014 allows local users to install or remove packages without authentication on Ubuntu, Debian, Fedora, and RockyLinux; fix is in PackageKit 1.3.5.</li>
<li><strong>ADT breach:</strong> ShinyHunters claims 10 million records stolen via a vishing attack that compromised an employee’s Okta SSO account and accessed ADT’s Salesforce instance; this is ADT’s third disclosed breach since August 2024.</li>
<li><strong>Bitwarden CLI supply chain attack:</strong> Bitwarden CLI 2026.4.0 was compromised via a malicious GitHub Action, stealing auth tokens and SSH keys.</li>
<li><strong>GPT-5.5 available on Vercel AI Gateway:</strong> GPT-5.5 and GPT-5.5 Pro are live, designed for long-running agentic work with better token efficiency than GPT-5.4; Xbow evals show missed vulnerabilities dropping to 10% in security testing.</li>
<li><strong>DeepSeek-V4:</strong> New open model with a hybrid attention architecture that reduces KV cache memory to roughly 2% of standard grouped query attention; V4-Pro-Max hits 80.6 on SWE Verified, placing it at parity with frontier closed models.</li>
<li><strong>Meta + AWS Graviton5:</strong> Meta signed a deal to deploy tens of millions of Graviton5 cores for agentic AI workloads, treating CPUs as the control plane for orchestration alongside GPU accelerators.</li>
<li><strong>Vault 2.0:</strong> HashiCorp released Vault 2.0 under IBM’s versioning lifecycle, adding Workload Identity Federation via OIDC tokens, SCIM 2.0 provisioning in beta, and SPIFFE JWT-SVID support; Azure authentication now requires explicit configuration.</li>
<li><strong>Zimbra XSS (CVE-2025-48700):</strong> Over 10,500 servers remain unpatched against an actively exploited XSS flaw that requires no user interaction beyond viewing a crafted email; CISA ordered federal agencies to patch within three days.</li>
<li><strong>AI phishing now top initial-access vector:</strong> Cisco Talos data shows phishing led 35% of Q1 2026 compromises, with AI-assisted campaigns achieving 54% clickthrough rates versus a 12% baseline and polymorphic campaigns averaging 1.8 emails before changing content.</li>
<li><strong>Cursor + Chainguard partnership:</strong> Cursor agents can now route dependency resolution to Chainguard’s verified artifact store of 2,300+ hardened container images instead of pulling from public registries like PyPI or npm.</li>
<li><strong>ClickFix macOS campaigns:</strong> Two separate campaigns — one linked to Lazarus Group targeting crypto and FinTech leadership via fake video calls, one using fake CAPTCHAs — are tricking users into running terminal commands that install credential stealers targeting 14 browsers and 200+ browser extensions.</li>
<li><strong>Vibe coding productivity study:</strong> A METR study found AI-assisted development actually increased task completion time by 19% despite developers believing it helped, alongside a 4x increase in code cloning and a drop in refactoring from 25% to roughly 3% of changes.</li>
<li><strong>FAST16 malware:</strong> SentinelOne researchers identified malware that may predate Stuxnet by five years, designed to corrupt floating-point calculations in engineering simulation software used in civil engineering and physics modeling.</li>
</ul>


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

Backend Digest

A developer news roundup covering several major stories: GitHub's data integrity incident drew criticism for downplaying impact on affected users. The Linux kernel deleted 138,161 lines of legacy code partly because AI-generated bug reports now consume 40% of maintainer time. Google is committing up to $40B to Anthropic as Claude's enterprise ARR hits $30B in 15 months. CrowdStrike's CEO warned of real-world cases where AI coding agents bypass security boundaries and steal credentials. Additional items include GlassWorm malicious VS Code extensions, a Cisco Firepower backdoor, GitHub App token format changes, a PackageKit privilege escalation CVE, the Bitwarden CLI supply chain attack, GPT-5.5 and DeepSeek-V4 releases, HashiCorp Vault 2.0, and a METR study finding AI-assisted development increased task completion time by 19%.

GitHub's data integrity mess, Linux kernel drops 138K lines to escape AI noise