Product updates, security insights, and technical deep-dives from our team. Learn how we're building the fastest security scanner for DevSecOps.
Most security tools give you a list of findings. They don't tell you which ones actually matter. A CVSS 9.8 in dead code and a CVSS 7.0 in your unauthenticated payment endpoint are not the same risk. Here's why the industry has this backwards — and where CodeSlick is going.
CodeSlick has 306 security checks — every one written against 2024-2025 attack patterns. We built CVE Coverage Guard: a weekly cron that monitors NVD, finds gaps in our own analyzer coverage, and opens GitHub PRs with Claude-generated test cases and check updates. First run: 254 CVEs, 2 real gaps found.
A transitive CVE slipped past CodeSlick because we were only reading package.json. Snyk caught it. 24 hours later we shipped full lockfile parsing, OSV batch queries, and a nightly dep monitor. CodeSlick now scans 1,083 packages per repo instead of ~50.
Vercel's April 2026 breach exposed customer environment variables that weren't flagged "sensitive." The taxonomy of sensitive vs. non-sensitive gives developers false confidence. Here's why every env var is an attack surface — and what CodeSlick detects before secrets ever reach a platform.
56 community MCP servers. March 18 → April 14. 100% have vulnerabilities. 89% have critical issues. 5,252 critical findings. Command injection in browser agents and Google's own gemini-cli. Scanned with CodeSlick v1.5.9.
August 2026 — the EU AI Act's high-risk AI provisions take effect. If your product ships AI-generated code in a regulated domain, you need a documented security audit trail today. What the law actually requires, and how to build it in four months.
AI coding agents start fresh every session — no memory of your repo's recurring SQL injections, hardcoded secrets, or worsening vulnerability trend. CodeSlick Security Memory changes that: your AI gets the full risk context before writing a single line of code.
Last week we audited 4 AI SDKs. Today we re-ran with 12 new MCP behavioral checks and added 4 more repos — CrewAI, MCP TypeScript SDK, Anthropic Python, Google Gemini JS. All original repos improved. The three structural failure modes are still in every codebase.
We ran CodeSlick across run-llama/LlamaIndexTS. Prototype pollution and command injection appear across the framework. The Postgres KV store has SQL injection in production code. Here is what the most widely used TypeScript RAG and agent library looks like from a security scanner.
Snyk Evo covers the security half of AI-assisted development. Nobody covers the maintainability half. We built both. Here is how CodeSlick and Endure together give teams the complete picture of their AI-assisted codebase.
We ran CodeSlick across supabase/supabase-js and supabase/auth-js — 2,410 findings. Weak random and hardcoded secrets appear across the default backend for AI-native applications. The uuid() generator uses Math.random(). GoTrueClient carries hardcoded secrets at 11 locations.
17 new checks for the vulnerability classes that live inside MCP tool handlers — prompt injection vectors, unauthorized financial API calls, system persistence writes, and unverifiable dependency execution. Covering what Snyk's Skill Inspector misses at the code level.
Your AI assistant follows your instructions — but does it follow your security policy? .codeslick.yml is a per-repo policy file that bans unsafe patterns, requires authorization checks, and controls pass/fail thresholds across your CLI pre-commit hook, GitHub App, and MCP Server. Define once, enforce everywhere.
916 vulnerabilities across microsoft/autogen and crewAIInc/crewAI — plus LangChain, Vercel AI, OpenAI, and MCP servers. AutoGen runs exec() on LLM-generated code by design. CrewAI has SQL injection in its memory layer. Here is what building on these frameworks actually means for your attack surface.
Run 308 security checks, detect secrets, scan dependencies, generate SBOMs, and inventory AI components directly from Cursor and Claude Desktop. No account required. All analysis runs locally. Includes generate_ai_bom — the only tool that inventories AI providers, models, and MCP tool registrations in a codebase.
We ran CodeSlick across vercel/ai, LangChain.js, openai-node, and MCP Servers — 5,381 files, 20,355 findings. Three failure modes appear in every repository: missing error handling, missing null checks, and hardcoded credentials in examples. Here is what the patterns mean for developers building on top of these libraries.
120,000+ combined GitHub stars. 97% had security findings. 75% had critical vulnerabilities. Command injection and SQL injection appear in agent-facing tools trusted to execute code on your machine. Here is what the MCP server ecosystem looks like from a security scanner's perspective.
Anthropic launched Code Review in Claude Code — a multi-agent PR reviewer that flags logic errors and costs $15–25 per scan. It's genuinely useful. Here's what it doesn't do, and why your security team still needs deterministic SAST.
MCP servers expose tools that AI models call with untrusted arguments. Command injection, path traversal, missing input validation — the same vulnerabilities we secured in web APIs are now hiding in your AI tooling. Here's what CodeSlick's 12 new MCP checks detect.
I ran Endure on the codeslick2 repo — the codebase that contains Endure — and found a calibration bug in the scorer. Here is what the analysis found, what changed, and what the tool still cannot do.
We ran CodeSlick on vercel/ai — 2,900 files, 1.5M weekly downloads — to understand what security debt looks like in a well-maintained project at scale. Prototype pollution, command injection, and AI-hallucinated methods hiding across eight packages.
Most systems don't collapse because of one bad decision. They weaken slowly — through intent drift, hidden coupling, and compounding opacity. Here's how fragility forms, and what structural inspection looks like.
AI handles 80% of predictable code remarkably well. The other 20% — rare inputs, legacy quirks, atypical integrations — is where systems fail. Here's the reasoning gap no model can fill, with real examples from Log4Shell, Heartbleed, and Facebook's 2021 outage.
Lost intent, silent drift, debt blindness, knowledge silos, maintenance roulette, stale code anxiety. Six real patterns — with six real incidents (Knight Capital, xz backdoor, CrowdStrike) — and how Endure addresses each one.
Anthropic just launched Claude Code Security — a research preview that finds complex vulnerabilities using semantic reasoning. The right question isn't whether to use it. It's where it fits in your pipeline. Here's the full picture.
The prevailing narrative says AI will automate cybersecurity and reduce headcount. It is structurally wrong. AI is expanding the attack surface, multiplying security roles, and creating a skills gap no automation can fill.
AI-accelerated development is widening a hidden gap: architectural entropy. When functions multiply at machine speed, structure decays at human speed. Here's the discipline that cannot be delegated.
41% of all code is now AI-generated. From agentic workflows to vibe coding—here's what's really happening and why governance is the new competitive advantage.
Comprehensive security audit of OpenClaw AI assistant reveals 277 CRITICAL vulnerabilities including command injection and SQL injection. Learn how to secure local-first AI platforms.
How we fixed two major problems: security issues that get forgotten, and "security passed" that still breaks production. Real improvements based on what developers kept telling us was broken.
We analyzed 10,000+ code snippets from GitHub Copilot, Cursor, and Claude Code. 47% contained security vulnerabilities. Most developers merged them without review.
The gap between human and AI reasoning is real—and it's showing up in your codebase. Learn how CodeSlick protects against AI-generated code threats.
Independent AI evaluation highlights CodeSlick's speed advantage and security-first approach. See how we compare to competitors.
Start securing your code with the fastest security scanner for GitHub teams.