# Tenki Blog > Product updates, guides, tutorials, and tips from the Tenki team. Tenki provides cloud infrastructure for code and agents: bare-metal GitHub Actions runners, an AI code reviewer for pull requests, and disposable Linux microVM sandboxes for AI coding agents. - Main site: https://tenki.cloud - Blog: https://tenki.cloud/blog - RSS (full content): https://tenki.cloud/blog/rss.xml - Markdown: append `.md` to any post URL below for the raw markdown version ## Blog Posts ### Security - [OIDC Keyless Auth for AWS, GCP, and Azure in GitHub Actions](): Configure OIDC federation for AWS, GCP, and Azure in GitHub Actions to eliminate long-lived cloud credentials from your CI pipelines. - [Cordyceps Exploits pull\_request\_target in 300+ Repos](): The Cordyceps CI/CD flaw lets anyone with a free GitHub account hijack privileged workflows at Microsoft, Google, and Apache. Checkout v7 doesn't fix it. - [38% of Workflows Are Vulnerable to Pwn Requests](): Datadog's 2026 DevSecOps report reveals 38% of orgs run workflows vulnerable to script injection. Here's what a pre-merge review gate catches. - [Claude Code Action Hijack: What Went Wrong](): A single GitHub issue could hijack any repo running Claude Code Action. Here's how the exploit chain worked and what to audit in your workflows now. - [AI Agent Prompt Injection in GitHub Actions](): Three confirmed incidents in six months prove AI coding agents in CI are a prompt injection target. Here's the threat model, the vulnerable patterns, and what pre-merge review catches. - [Miasma Worm Scraped CI Runner Memory for Secrets](): The Miasma/Hades supply chain worm compromised 73 Azure repositories by scraping CI runner memory for cloud secrets. Here's what a pre-merge review layer catches before the worm gets in. - [GitHub OIDC for Dependabot: What It Secures, What It Doesn't](): GitHub's expanded OIDC eliminates long-lived secrets from Dependabot workflows. But OIDC secures the token, not the PR. Here's what's still exposed. - [Axios npm Supply Chain Attack: What CI/CD Teams Must Lock Down](): The Axios npm compromise planted a North Korean RAT via post-install hooks. Here's how it hit CI/CD pipelines and the lockdown checklist your team needs now. - [GitHub's Internal Repo Breach and Your CI Trust Boundary](): GitHub disclosed unauthorized access to internal repos. Here's why your CI security needs a review gate outside the platform's trust boundary. - [GitHub MCP Scanning Is Here. Your Review Layer Still Matters.](): GitHub shipped MCP-native secret and dependency scanning, but tool-level checks inside the agent loop can't replace a pre-execution review gate. - [GitHub Connects Code to Cloud Risk via Defender](): GitHub's new Defender for Cloud integration maps code flaws to running infrastructure. Here's why the CI runner is where you actually stop them. - [Prompt Injection in AI-Powered GitHub Actions](): AI agents that triage issues, review PRs, and label code introduce a prompt injection attack surface that traditional workflow hardening completely misses. - [The PR Comment That Hijacked Three AI Agents](): A single prompt injection pattern exfiltrated API keys from Claude Code, Gemini CLI, and GitHub Copilot simultaneously. Here's why it worked and what to audit now. - [OWASP AI Agent Security Top 10: CI/CD Audit Guide](): Map each OWASP ASI risk category to concrete CI/CD controls, from token scoping and action pinning to agent behavioral monitoring. - [CI/CD Policy as Code: From YAML Security to Centralized Governance](): Per-workflow YAML security doesn't scale. Here's how centralized, declarative policies transform CI/CD security from scattered configs into enforceable governance. - [When Your Security Scanner Gets Compromised](): The Trivy supply chain compromise shows that CI/CD security tools themselves are high-value targets. Here's what happened and how to build pipelines that don't trust any single tool. - [GitHub Actions Workflow Lockfiles Are Coming](): GitHub's new dependencies: section in workflow YAML locks direct and transitive action dependencies by commit SHA, making CI builds deterministic and reviewable. - [GitHub Actions Artifact Attestations: SLSA Provenance and Supply Chain Defaults](): GitHub's artifact attestation feature uses Sigstore to generate SLSA provenance for your builds. Here's how it works, what it proves, and what changes as defaults tighten. - [GitHub Actions Permissions: Lock Down GITHUB\_TOKEN](): Default GITHUB\_TOKEN permissions are dangerously broad. Here's how to scope every job to minimum access and eliminate static cloud credentials with OIDC. ### Code Review - [Human Review of AI Code Is a Structural Mismatch](): At 97% AI tool adoption, expecting developers to catch subtle errors in AI-generated code through manual review isn't a skills gap. It's a cognitive model built for a world that no longer exists. - [AI Code Review Tools Compared: Bugbot, CodeRabbit, Graphite, Copilot](): Four AI code review tools, four different approaches. Here's which one actually fits your team's workflow and which ones create more noise than they eliminate. - [Your Tests Pass. Your AI Agent's Code Is Still Wrong.](): Test suites catch regressions developers anticipated. AI agents introduce a class of bugs tests were never designed to find. Here's the missing CI layer. - [Each AI Agent Has a Distinct Bug Fingerprint](): Greptile's data from 27.6% AI-generated PRs shows Claude, Codex, and Cursor fail in categorically different ways. Here's how to configure review rules that match. - [The CI Runner Is Where AI Review Gets Teeth](): Comment-based AI code review tools can be dismissed with a click. A required status check on a CI runner can't. Here's why that distinction matters. - [Add Tenki AI Code Review to GitHub in Under 30 Minutes](): Install the Tenki Code Reviewer GitHub App, interpret your first flagged PR, tune severity thresholds, and enforce reviews as a merge gate. - [97% AI Coding Tool Adoption: What It Means for Your Merge Gate](): Black Duck's June 2026 study found 97% of developers use AI coding tools. Your pre-merge review gate needs to catch what those tools miss. - [Copilot Code Review Now Pulls from MCP Servers](): GitHub's June 2 changelog adds MCP server connections to Copilot code review. That means external servers now shape what gets flagged at your merge gate. - [Your AI Coding Agent and Code Reviewer Shouldn't Be the Same Company](): Cursor now owns Graphite. When the same vendor writes your code and reviews it, the quality signal gets weaker. Here's why independent review matters. - [Copilot Code Review Now Eats Your Actions Minutes](): GitHub Copilot code review now burns Actions minutes on every PR. Here's the billing math and why Tenki's per-review pricing sidesteps the problem. - [Graphite Is Now a Cursor Product: Why Independent Code Review Matters](): Cursor acquired Graphite, merging AI code generation and review under one roof. Here's why Tenki keeps its review engine independent from any coding toolchain. - [Tenki vs Graphite: AI Code Review Pricing](): Graphite charges $40/user/month for unlimited AI reviews inside a full PR platform. Tenki charges $1/review as a CI merge gate. Here's how to choose. - [GitHub Code Quality GA: What $10/Committer Blocks](): GitHub Code Quality goes paid July 20 at $10/committer/month. Here's what the quality gate enforces, what it misses, and where semantic review still matters. - [CI Green Is Not Merge Safe](): A passing CI build proves your code compiles and existing tests pass. It says nothing about security, logic correctness, or design quality. Here's why that distinction matters. - [GitHub Code Quality Is a Gate. It's Not Independent.](): GitHub Code Quality shares billing, org settings, and budget caps with your repos. That makes it structurally different from a gate living in CI YAML. - [SpaceX Bought Cursor: Your Review Gate Can't Live There](): SpaceX's $60B Cursor acquisition puts code generation and review under one defense contractor. Here's why your merge gate needs structural independence. - [Stop Counting Comments: Action Rate Is the AI Code Review Metric That Matters](): Comment volume is a vanity metric. The percentage of AI review comments that actually change code before merge is the only signal that predicts developer trust. - [GitHub PR Coverage Is Useful. Here's What It Can't Tell You.](): GitHub's new PR code coverage feature shows what your tests exercise. Tenki's review layer shows what the exercised code actually does wrong. - [GitHub Made Code Review a Metered Utility](): GitHub Copilot code review now bills per token and per Actions minute. When quality gates have variable costs, teams optimize spend over depth. - [GitHub Measures Copilot Adoption. Tenki Measures What Passes Review.](): GitHub's new cohort metrics track which developers use Copilot. That's adoption data, not quality data. The metric that matters is catch rate, and it lives in your CI logs. - [GitHub Gets Agent PR Review Right, Not the Fix](): GitHub's agent PR review guide nails the problem. But checklists don't scale. Automated review gates must filter agent output before humans touch it. - [Copilot Now Batch-Fixes Its Own Reviews. Here's the Gate It Skips.](): GitHub's Fix batch with Copilot closes the AI review-to-fix loop. Without an independent CI gate, that loop runs unchecked. - [Feedback Sensors for Coding Agents: Wiring Quality Gates Into Self-Correction Loops](): Compilers, linters, and test suites wired directly into coding agent workflows catch more defects than post-hoc review alone, turning code review into a verification step. - [Tenki catches 2x more real bugs than any other AI reviewer. Here's the benchmark.](): We ran Tenki against six other AI code review tools on 50 real bug-introducing pull requests from cal.com, Sentry, Grafana, Keycloak, and Discourse. Here is how the benchmark works, why F1 is the right metric, and what the results mean. - [Signed Agent Commits Prove Authorship, Not Correctness](): GitHub's Copilot cloud agent now signs its commits, but a verified badge proves who wrote the code — not that the code is safe to merge without review. - [Claude Code Review: What $15 Per PR Means for Enterprise Budgets](): Anthropic's multi-agent Code Review costs $15-25 per PR. Here's the cost model for teams running 50+ PRs a day, and what it reveals about the economics of automated review. - [Tenki vs Greptile: Review Depth, Noise, and Cost](): Greptile indexes your entire codebase for cross-file context. Tenki focuses on actionable, merge-blocking signal. Here's how they compare on review quality, false positives, and pricing. - [Tenki vs Ellipsis AI Code Review: Scope, Fixes, Pricing](): Tenki and Ellipsis both review pull requests with AI, but they differ on what they catch, how they fix issues, and what they cost. Here's a direct comparison. - [AI Reviewing AI: Shared Blind Spots in AI-on-AI Code Review](): When AI agents generate code and AI tools review it, shared training patterns create systematic blind spots. Here's how to break the homogeneity trap. - [What to Look For When Reviewing Agent-Generated Code](): AI agents write syntactically clean code that can be semantically wrong. Here's a practical checklist for catching the failure modes humans miss in agent-authored PRs. - [Reviewing AI-Generated Code: Different PRs Need Different Patterns](): AI coding assistants produce structurally different code than humans. Your review process needs to adapt or you'll miss the bugs that matter most. - [Measuring AI Code Review ROI the Right Way](): Most teams adopt AI code review tools without baselines, making ROI impossible to measure. Here's a framework for tracking defect escape rates, review time, and the metrics that actually matter. - [PR Review and CI Are Two Different Systems](): CI validates what the code does. Review validates what the code means. Conflating them is why bugs pass both gates undetected. - [AI Made Writing Code Cheap. Judgment Didn't Scale.](): AI writes code faster than ever, but review quality is falling behind. The gap between velocity and judgment is the real risk most teams haven't named yet. ### AI Agents - [Cap What Copilot Spends, Not What It Ships](): GitHub's new AI credit pools and session limits cap token costs. They don't cap what reaches main. Here's why you need both. - [Agentic CI Solved Throughput. The Bottleneck Is the Gate.](): GitHub solved CI throughput with parallel steps and agentic workflows. The bottleneck now is whether your merge gate can tell safe PRs from broken ones. - [GitHub Dropped PAT Auth for Agents. Your Merge Gate Just Got More Important.](): GitHub's GITHUB\_TOKEN now powers agentic workflows without PATs. That removes the last credential barrier to automated PRs, making your merge gate the final checkpoint. - [Copilot Sandboxes Are Live. Your Merge Gate Still Matters.](): GitHub Copilot sandboxes isolate agent execution, but isolation doesn't prove code is safe to merge. Here's where a pre-merge review gate fits in. - [GitHub Agentic Workflows Need a Review Gate](): GitHub Agentic Workflows let CI jobs spawn AI agents that commit code and open PRs. The output still needs an independent review gate before it merges. - [Namespace Runs the Agent. Tenki Reviews What It Wrote.](): Namespace Devboxes give Claude agents isolated compute. Tenki reviews the PR diff at the CI boundary before it merges, regardless of where the agent ran. - [Copilot Token Billing Is Coming: What Enterprise Teams Need Now](): GitHub's sign-up freeze and shift to token-based billing hit June 1. Here's how enterprise teams can model token consumption and avoid budget surprises. - [The Composable AI Coding Stack Is Here](): Cursor, Claude Code, and Codex aren't competing anymore. They're becoming orchestration, execution, and review layers in a stack teams should architect deliberately. - [Copilot's April 24 Data Policy: What Changes and How to Opt Out](): GitHub's April 24 policy update enables Copilot interaction data training by default for Free, Pro, and Pro+ users. Here's what's affected and how to opt out. - [AI Code Creates 1.7x More Review Issues](): CodeRabbit's data shows AI-generated PRs produce 1.7x more issues than human code. The fix isn't better prompts — it's review architecture built for agent-written code. - [AI Coding Tools: Creating Senior Engineers or Accelerating Them?](): GitClear's 2026 research shows AI power users produce 4-10x more code, but the data reveals selection bias, not skill creation. Here's what engineering leaders should measure instead. - [MCP Security Scanning: Audit Your AI Agent's Tools](): MCP tool servers expand your AI agent's attack surface with every integration. Here's how to audit tool descriptions, detect poisoning attacks, and build an allowlist before they become vectors. - [What Stripe's 1,300 Agent PRs Per Week Reveal About CI at Scale](): Autonomous coding agents are flooding CI pipelines with pull requests. Stripe's Minions system shows what breaks first and how to plan for it. - [The Hidden CI Tax of AI Coding Agents](): AI coding agents generate 3-5x more PRs per developer, and every one triggers your CI pipeline. Here's how to quantify the cost and control it. - [Claude Code Source Leak: What It Reveals About AI Agent Trust](): The leaked Claude Code source exposes fake tools, frustration-detection regexes, and an undercover mode. Here's what engineering teams should demand from AI coding agents. - [Audit AI Agent Behavior in CI with Session Traces](): GitHub's new Actions run summary configs and commit-to-session tracing create a real audit trail for AI agents in CI. Here's how to build observability around it. - [Copilot Coding Agent Observability: From Black Box to Audit Trail](): GitHub shipped five observability features for Copilot coding agent in one week. Here is what commit tracing, validation controls, and live monitoring mean for teams adopting AI agents. - [GitHub Agentic Workflows Explained](): GitHub Agentic Workflows replace YAML with Markdown and hand execution to an AI agent. Here's the security model, tradeoffs, and whether to adopt now. ### GitHub Actions - [GitHub Actions Cost Optimization in 2026](): A platform engineer's guide to cutting GitHub Actions spend through runner right-sizing, cache strategies, concurrency controls, and provider alternatives. - [The AI Agent PR Playbook: GitHub Actions, Branch Policies, and Review Thresholds](): A reference guide for configuring GitHub Actions workflows, branch protection rulesets, and review thresholds when most of your PRs come from AI coding agents. - [Jenkins to GitHub Actions Migration Guide](): Translate your Jenkinsfile to GitHub Actions YAML, convert parallel stages and shared libraries, migrate credentials to OIDC, and add Tenki review from day one. - [GitHub Actions Required Workflows: Org-Wide CI Gates](): Use GitHub organization rulesets with required workflows to enforce consistent CI gates across every repository, no per-repo setup needed. - [GitHub Actions Runner Images in 2026: The Complete Selection Guide](): Three major runner image migrations hit in June 2026. Here's how to pick the right base image for Ubuntu, Windows, macOS, and ARM64 workflows. - [GitHub Is Building the Full CI Stack. Keep Your Review Layer Independent.](): GitHub shipped five Copilot CI integrations in one week. When one vendor owns your executor, reviewer, and scanner, a single regression hits everything. - [Migrate GitHub Actions to Node.js 24 Before the Deadline](): Node.js 24 becomes the default for JavaScript-based GitHub Actions on June 16. Node 20 disappears from runners entirely in fall 2026. Here's how to audit, test, and migrate safely. - [Flaky Test Quarantine in GitHub Actions](): Stop rerunning failed jobs and hoping for green. Build a quarantine pipeline that auto-detects flaky tests, isolates them from PR checks, and retires them before they rot. - [GitHub Actions Caching: npm, Docker, and Turborepo](): A practical guide to every caching layer in GitHub Actions, from lockfile-based npm keys to Docker layer caching and Turborepo remote cache. - [Lint Your GitHub Actions Workflows with actionlint](): actionlint and OpenSSF Scorecards catch expression injection, deprecated syntax, and misconfigurations in your workflow YAML before they reach production. - [Monorepo CI with GitHub Actions: Selective Builds Done Right](): Naive path filters miss cross-package dependencies. Dependency-aware selective builds with Turborepo, Nx, and conditional job execution fix monorepo CI without over-building or under-building. - [GitHub Actions Cost Optimization: Cut Your Billable Minutes by 40–70%](): Most teams treat their GitHub Actions bill as a fixed cost. It isn't. Path filters, concurrency cancellation, job splitting, and runner selection routinely cut billable minutes by 40–70% without sacrificing test coverage. Here's how to audit your workflows and find the savings. - [GitHub Merge Queue in 2026: How It Works \& Handling Flaky Required Status Checks](): GitHub's merge queue prevents broken merges on busy branches, but common gaps trip up most teams who enable it. Here's how the queue actually works, how to configure it properly, and how to stop flaky tests from blocking your entire pipeline. - [Custom GitHub Actions: JavaScript, Docker, and Composite](): JavaScript actions start fast and run everywhere, Docker actions control the full environment, and composite actions glue shell steps together. Here's how to choose. - [Build and Push Docker Images with GitHub Actions](): A complete guide to building, tagging, caching, and pushing Docker images to GHCR, ECR, and Docker Hub from GitHub Actions workflows. - [Release Automation with GitHub Actions](): Automate version bumps, changelogs, and publishing with semantic-release, Changesets, or release-please on GitHub Actions. - [Reusable Workflows vs. Composite Actions: Pick Right](): Reusable workflows and composite actions both eliminate copy-pasted YAML, but choosing the wrong one creates tech debt that's painful to reverse at scale. ### Runners - [Self-Hosted Runners in 2026: ARC, Security, Cost](): Everything platform engineers need to know about self-hosted GitHub Actions runners: ARC autoscaling, ephemeral modes, security hardening, and where managed runners eliminate the operational burden. - [GitHub Actions Runner Enforcement Starts June 29](): Brownouts for self-hosted runners below v2.329.0 begin June 29. Here's the full timeline, how to audit your fleet, and why managed runners sidestep the problem entirely. - [Self-Hosted Runners Are a Maintenance Trap](): Self-hosted GitHub Actions runners look cheaper on paper. Once you factor in fleet management, patching, and on-call, managed runners win for most teams. - [Tenki vs. GitHub-Hosted Runners: Pricing, Specs, and What Ships with Your Build Minutes](): Tenki runners cost $0.002/core/minute on x64, undercutting GitHub-hosted runners by up to 33%, and bundle AI code review at the PR boundary for $1/review. - [CI Vendor Shutdowns Are a Pattern, Not a Fluke](): BuildJet and Cirrus CI both shut down in 2026. Here's what their exits reveal about managed runner platform risk and how to evaluate your next CI vendor. - [Cut Your GitHub Actions Bill by 90% in 2026](): A practical playbook for slashing CI/CD costs: caching, parallelization, runner selection, and self-hosted alternatives that actually deliver savings. - [Copilot Code Review Now Costs Actions Minutes](): GitHub's June 1 billing change adds Actions minutes to every Copilot code review on private repos. Here's what it costs and how self-hosted runners cut the bill. - [Depot CI Says It Rebuilt CI for Agents. Here's Tenki's Take.](): Depot rebuilt CI from scratch for the agentic era. Tenki delivers the same speed gain with a one-line YAML change and no migration. - [GitHub Actions Runner Pricing Postponed: Model Your CI Costs Now](): GitHub shelved the $0.002/min self-hosted runner charge after community backlash. Here's how to model your actual CI costs before revised pricing arrives. - [Migrate BuildJet to Tenki Runners: Labels, Cache, Costs](): BuildJet shut down March 31, 2026. Here's the concrete migration path to Tenki Runners, with label mappings, cache changes, cost comparisons, and YAML diffs. - [GitHub Actions Runner Showdown 2026](): Tenki vs. WarpBuild vs. Blacksmith vs. Namespace: real pricing breakdowns, benchmark data, and a migration guide for teams re-evaluating CI runners. - [ubuntu-latest is Ubuntu 24.04: What to Audit](): GitHub's ubuntu-latest label now resolves to Ubuntu 24.04, removing Python 2, shipping OpenSSL 3, and changing default tool versions across runner images. ## Categories - [Security](): Supply chain security, SLSA, OIDC, secrets management, policy as code, and hardening. - [Runners](): Self-hosted runners, GPU runners, pricing, billing, and cost optimization. - [AI Agents](): AI coding agents, Copilot, agentic workflows, and LLM-powered automation. - [Code Review](): Code review tools, code quality, productivity measurement, and ROI. - [GitHub Actions](): Core GitHub Actions features, workflows, caching, custom actions, migrations, and tooling. - [Sandbox]()