Tech Pulse — Mar 28, 2026

Tenable: Full March 2026 Patch Tuesday analysis — CVE-2026-21262 & CVE-2026-26127 →

DeepMind AlphaEvolve: AI Code Agent Recovering 0.7% of Google's Global Compute

Google DeepMind has revealed that AlphaEvolve — a Gemini-powered autonomous coding agent — has been running continuously inside Google's own infrastructure for over a year, making it the first confirmed large-scale AI agent with verified ROI at hyperscaler scope. The system uses evolutionary search to discover novel algorithms and rewrite production code.

• Compute recovery: AlphaEvolve continuously reclaims 0.7% of Google's worldwide computing resources by optimizing scheduling and kernel-level operations — at Google's scale, that represents hundreds of millions in annual cost avoidance.
• Training acceleration: A key Gemini training kernel was sped up by 23% after AlphaEvolve autonomously rewrote the underlying algorithm — a breakthrough that manual optimization had not achieved.
• Scientific expansion: AlphaEvolve's accelerated access program has been extended to U.S. Department of Energy National Laboratories for scientific algorithm design, including computational physics and climate modeling.
• Complexity theory: The agent has solved open problems in algorithmic complexity, advancing the frontier of what automated code discovery can achieve beyond existing human-written solutions.

DeepMind Blog: AlphaEvolve — evolutionary code agent at production scale →

GitHub: New PR Dashboard, Credential Revocation API & Actions Security Roadmap

GitHub shipped three significant developer-facing updates this week, addressing long-standing pain points in pull request management, supply-chain credential exposure, and CI/CD pipeline security following a wave of tj-actions and trivy-action supply-chain attacks.

• PR Dashboard public preview: The refreshed github.com/pulls introduces a PR inbox and saved views — enabling teams to triage, filter, and prioritize open reviews across multiple repositories from a single pane.
• Credential Revocation API: Programmatic revocation of GitHub OAuth tokens and GitHub App credentials — any exposed token found in code, logs, or third-party services can now be revoked via API call without manual portal access.
• Actions 2026 Security Roadmap: GitHub published its full Actions security strategy in direct response to the tj-actions/changed-files, Nx, and trivy-action supply-chain compromises — including mandatory workflow pinning enforcement and provenance attestation requirements.
• Copilot Coding Agent metrics: GitHub now separately tracks Copilot Coding Agent (CCA) users in usage dashboards — distinguishing agentic multi-step task completion from single-line IDE completions.

GitHub Changelog: New PR Dashboard in public preview →

Nscale $2B Series C — Europe's Largest-Ever AI Infrastructure Raise

UK-based AI hyperscaler Nscale has closed a $2 billion Series C — the largest venture capital round in European history — valuing the company at $14.6 billion. The raise consolidates Nscale's joint venture with Aker ASA and positions it as Europe's primary sovereign AI infrastructure alternative to US hyperscalers.

• GPU deployment: Plans to deploy approximately 200,000 NVIDIA GB300 GPUs across data centers in Europe and the US — creating one of the highest-density AI training clusters outside the US hyperscalers.
• Strategic investors: Consortium led by Aker ASA and 8090 Industries; co-investors include NVIDIA, Dell, Nokia, Goldman Sachs, Citadel, Jane Street, Point72, and Lenovo — unusually broad institutional and strategic backing.
• UK commitment: £2 billion committed to UK data center buildout by 2028, addressing European sovereign AI compute demand driven by GDPR and EU AI Act compliance requirements.
• Board expansion: Sheryl Sandberg, Susan Decker, and Nick Clegg join as new board members — signaling governance maturity ahead of potential future public offering.

Nscale: Full $2B Series C press release →

Shield AI $1.5B Series G — Defense Autonomy at $12.7B Valuation

San Diego-based Shield AI has raised a $1.5 billion Series G at a $12.7 billion valuation, cementing its position as the leading AI autonomy stack for military drones and aircraft. The round represents the largest defense-AI raise of Q1 2026 and signals that autonomous weapons systems are moving from R&D into large-scale production deployment.

• Hivemind stack: Shield AI's core product enables drones and aircraft to operate autonomously without GPS, communications links, or human pilots — critical for contested electromagnetic environments.
• Investors: Round led by Advent International and JPMorgan Chase's Strategic Investment Group — both with deep aerospace and defense portfolio experience.
• Acquisition target: Plans to use capital to acquire Aechelon Technology, a defense simulation specialist, consolidating software, autonomy, and synthetic training environments into one platform.
• Market context: AI defense companies are consolidating — full-stack government platforms integrating autonomy, simulation, and live fleet management are becoming the new standard contract requirement.

Crunchbase News: Shield AI $1.5B Series G details →

OpenAI GPT-5.4 Codex Plugins + SoftBank's $40B Bridge Loan

OpenAI's GPT-5.4 ships with a 1-million-token context window and autonomous multi-step workflow execution across desktop environments — a direct upgrade path for developers building agentic applications. Separately, SoftBank has secured a $40 billion bridge loan to fund further OpenAI investment, underscoring the scale of capital committed to the frontier AI race.

• 1M context window: GPT-5.4 supports up to 1 million tokens in a single context — enabling full codebase ingestion, long document analysis, and sustained multi-session agentic workflows without summarization truncation.
• Codex plugins: Standardized repeatable AI workflows with 20+ initial integrations including Figma, Notion, Gmail, and Slack — enabling production-grade automation without custom API glue code.
• OSWorld-V benchmark: GPT-5.4 scores 75% on OSWorld-V (simulated desktop productivity tasks) — highest score on autonomous software execution benchmarks to date.
• SoftBank $40B: Arranged with JPMorgan, Goldman Sachs, Mizuho, SMBC, and MUFG, running to March 2027 — total SoftBank commitment to OpenAI now exceeds $50 billion, the single largest position in the AI stack.

AI Model Releases 2026: GPT-5.4 technical details and benchmarks →

Meta & Alphabet Found Liable in Child Harm Trials — Legal Precedent Set

A Los Angeles jury has returned the first US verdict holding both Meta and Alphabet (Google) liable in child harm lawsuits, ordering a combined $6 million in damages. While the dollar amount is modest, the verdict establishes legal precedent that platform recommendation algorithms can carry product liability — a ruling with profound implications for how developers and product teams design content distribution systems.

• Precedent value: This is the first time a US jury has found Big Tech liable for algorithmic harm to minors — hundreds of similar pending cases will now use this ruling as a template for damages claims.
• Algorithm design risk: The verdict targets recommendation engine behavior, not just content moderation failures — suggesting courts are prepared to treat algorithmic amplification as a design defect subject to product liability law.
• Regulatory convergence: Ruling aligns with the EU Digital Services Act's algorithmic accountability requirements and KOSA (Kids Online Safety Act) provisions currently advancing in US Congress.
• Developer impact: Teams building recommendation systems, social feeds, or content ranking algorithms should now treat algorithmic amplification of harmful content as a legal liability risk, not just a moderation concern.

TechStartups: Meta & Alphabet child harm trial verdict — full coverage →