Expert insights on AI code quality, security, and local-first development
The Miasma worm planted .claude/settings.json and .gemini/settings.json files in 73 Microsoft GitHub repositories to harvest cloud credentials the moment a developer opened the project. Here is the full attack anatomy and the local-first checks that stop it.
Read full article →Pillar Security researchers chained a prompt injection in a public GitHub issue to a full supply chain compromise of the gemini-cli repository, earning a CVSS 10 score. Here is the four-step attack, the root cause, and what local-first quality gates can enforce before your own agentic workflows reach the same fate.
Read full article →A fed-up Java library maintainer embedded hidden AI agent instructions in jqwik to delete vibe coders' output. The package passed all CVE scans. Here is the new threat model and what your pipeline should check that Dependabot cannot.
Read full article →A bootstrap API and a live feature-flag sync mean Anthropic can modify what Claude Code is told to do, at startup and every 60 seconds after, without shipping an update. Here is what the mechanism looks like, what the source-code leak revealed, and why local-first validation matters more than ever.
Read full article →The NSA published a formal MCP security advisory today. Here is what the advisory gets right, what it misses, and five concrete steps every AI developer should take this week.
Read full article →A new paper from arxiv shows LLM coding agents lose 30+ accuracy points as structural constraints accumulate, a phenomenon called constraint decay. Here is what it means for teams shipping AI-generated backend code and the quality gates that catch what tests miss.
Read full article →When Claude Code Agent Teams or Cursor 3.0 orchestrate sub-agents, a single poisoned tool description or shared memory entry propagates through the entire pipeline. Here is how transitive prompt injection works in multi-agent coding workflows and what local gates stop it before your orchestrator delegates to the first sub-agent.
Read full article →When AI coding agents hallucinate package names, attackers register those exact names with malicious payloads. Here is the complete attacker playbook for slopsquatting: how they identify targets, register packages, and why AI agents make better victims than humans ever were.
Read full article →AI-generated PRs arrive 5x faster than human ones, but review speed has not changed. When reviewers are drowning in backlog, approval becomes a formality. Here is the concrete list of defects that automated quality gates catch that humans consistently miss under cognitive load.
Read full article →The Clinejection attack turned Cline's own GitHub Actions bot into a supply chain weapon, installing rogue agents on 4,000 developer machines. Here is the anatomy of the attack and the local gates that stop the next one.
Read full article →The Mini Shai-Hulud worm compromised 84 @tanstack packages in six minutes, bypassing valid OIDC provenance and 2FA. Here is exactly what LucidShark SCA surfaces in your editor before the malicious payload runs.
Read full article →Every Claude Code session loads CLAUDE.md and workspace settings without integrity verification. Deeplink attacks, malicious PRs, and supply chain modifications can inject persistent behavioral instructions. Here is how to treat your agent config as the security boundary it is.
Read full article →Claude Code, Codex CLI, and Gemini-CLI share a TOCTOU trust model that lets a single malicious commit permanently own every future agent session. All three vendors closed the report as informational. Here is what your team can do instead.
Read full article →Practical techniques for auditing overnight AI agent commits before they reach production, covering diff analysis, semantic review, security checks, and automated gates.
Read full article →Georgia Tech tracked 35 CVEs filed in March 2026, with 27 traced to Claude Code output. Here is what the data actually means for security teams and why volume acceleration changes the entire risk calculus for AI-assisted development.
Read full article →GitHub Copilot's silent co-author attribution in VS Code commits sparked outrage, but the real risk is not in the git log. It is in how co-author tags change code review behavior and what automated quality checks exist for the code behind them.
Read full article →CVE-2026-26268 turns a routine git clone into remote code execution by embedding a malicious bare repository with a pre-commit hook that Cursor AI agent triggers automatically. Here is how the attack works and what local gates stop it.
Read full article →Socket Research Team caught four SAP CAP npm packages carrying malicious preinstall hooks on April 29, 2026. Combined weekly downloads: 572,000. The script downloaded and executed a Bun binary. Here is what happened, why AI coding agents make this worse, and exactly what to do right now.
Read full article →Anthropic confirmed the MCP STDIO remote code execution vulnerability is by-design behavior. With 150 million downloads affected and no protocol fix coming, here is the concrete checklist every team running agentic workflows needs to implement this week.
Read full article →AI coding agents invent package names that do not exist. Attackers register those exact names with malicious payloads. Here is how to validate every AI-generated import before it reaches your lockfile.
Read full article →Anthropic removed Claude Code from its Pro plan, banned organizations without warning, and went down three times in April. Here is what your quality pipeline should look like when the AI layer is unavailable.
Read full article →AI coding agents excel at writing code that satisfies your test suite and at writing tests that game coverage metrics. Here is why your coverage numbers are probably lying to you, and the pre-commit check that catches it.
Read full article →Anthropic's Project Glasswing found 35 CVEs in March 2026 alone, up from 6 in January. If AI can find zero-days at this velocity, AI-generated code is the most attractive attack surface on the internet. Here is the local quality gate that changes your risk profile.
Read full article →The OWASP Top 10 for Agentic Applications reframes AI security around autonomous action. Here is what each risk means for teams shipping code with Claude Code and agentic workflows.
Read full article →A maliciously crafted Git branch name with hidden Unicode characters could silently exfiltrate your GitHub access token via OpenAI Codex. Here is how the attack works, how to detect it, and what the pattern means for AI coding tool security.
Read full article →The TeamPCP group hijacked 75 tags of trivy-action on GitHub, turning every CI/CD pipeline that called Trivy into a silent credentials exfiltration machine. Here is what happened, why it worked, and how to harden your scanner setup before the next campaign.
Read full article →The Axios supply chain attack exposed a critical gap: provenance attestations existed but a legacy token bypassed them entirely. Here is how to close that gap with npm provenance, SLSA level 2, and automated SCA checks.
Read full article →The axios npm supply chain attack of March 2026 deployed a RAT through AI coding agents autonomously running npm install. Learn how MCP connector poisoning works and how local SCA scanning stops it before it reaches your codebase.
Read full article →AI-generated code is flooding SAST pipelines with false positives — Ghost Security found a 91% false positive rate in open source repositories. Learn how hybrid LLM+SAST analysis and local-first tooling cuts noise without missing real threats.
Read full article →A new class of supply chain attack targets AI coding agents directly. Learn how malicious npm and PyPI packages inject instructions into Claude Code sessions — and how local-first SCA scanning stops them.
Read full article →AI coding agents like Claude Code are prolific duplicators. They copy logic, repeat utility functions, and re-declare constants across files without hesitation. Here is why that compounds into real technical debt, and how LucidShark's duplication analysis catches it before it ships.
Read full article →RSAC 2026 researchers disclosed CVEs across 100% of tested AI IDEs including Claude Code, Cursor, and Copilot. Here's what the findings mean for teams shipping AI-generated code, and why a local quality gate beats a $25-per-review cloud service.
Read full article →A post-mortem on the fastest database breach of 2026 and the quality gate that would have stopped it cold. Learn how LucidShark's security checks would have caught the hardcoded API keys, disabled RLS, and misconfigurations that led to the complete exposure of Moltbook's database in under three minutes.
Read full article →Two critical CVEs in Claude Code expose a new threat model for AI coding tools. Learn what CVE-2025-59536 and CVE-2026-21852 reveal about reviewing AI-generated code and why you need local security gates.
Read full article →AI code generation is transforming development, but vibe coding comes with security risks. Learn the essential checklist every developer needs to validate AI-generated code before shipping to production.
Read full article →Discover why local-first code quality tools are essential for AI-assisted development with Claude Code. Privacy, speed, and better integration.
Read full article →Comprehensive comparison of AI code review tools. See which tools actually catch bugs in real-world codebases and how they integrate with your workflow.
Read full article →