Skip to content

AI Agent GitHub Digest — 2026-11-19

Nov 19, 20261 min
TL;DRPydantic AI v2.70.0 fixes two high-severity ReDoS bugs, and CrewAI 1.15.42 patches a soupsieve CVE; Claude Code's red-team skill pack Claude-BugHunter breaks 1,000 stars in a day, Google wires Agent Skills into its own design tool Stitch, and Mastra ships its own issue-in, PR-out Factory template.

Look up & save

Loading dictionary…

中文版

Today's Highlight

Today's thread is security catching up with how fast these frameworks keep growing. Pydantic AI fixed two high-severity ReDoS bugs hiding in its own regexes, and CrewAI patched a CVE in a third-party dependency on the same day. On the tooling side, Mastra followed yesterday's Tencent LoopForge with its own issue-in, PR-out Factory template, and Google wired the Agent Skills open standard into its own design tool, Stitch — skills-as-plugins is turning into standard equipment across product lines.

elementalsouls/Claude-BugHunter 11,028 (+1,005)

GitHub · Python · MIT

  • What it is: A red-team/bug-hunting skill pack for Claude Code — 83 skills and 15 slash commands, curated from 681 disclosed HackerOne reports across 24 vulnerability classes, plus attack-chain templates for enterprise identity systems (M365/Entra, Okta, vCenter).
  • Why it matters: Once installed, Claude Code doesn't just answer security questions — it works through a four-layer skill chain (recon → classify vulnerability → validate → write report) that triggers automatically from what you describe, with built-in Burp MCP integration and report templates. It packages a bug hunter's actual workflow into a reusable install, not a one-off prompt. The +1,005 stars in a day makes it one of today's fastest-growing agent-related repos.
  • Tech stack: Claude Code Skills + 83 skill files + Burp MCP + the cbh CLI (installed via pipx)
  • Getting started: Easy — one line, /plugin install claude-bughunter@elementalsouls, and every skill auto-triggers from what you describe, no manual invocation needed

google-labs-code/stitch-skills 10,350

GitHub · TypeScript · Apache-2.0

  • What it is: Google Labs' own Agent Skills library, built to pair with Stitch (Google's AI UI design tool) and its MCP server, following the Agent Skills open standard — installable into Codex, Antigravity, Gemini CLI, Claude Code, and Cursor.
  • Why it matters: This isn't a community plugin — Google split "design," "build," and "utility" skills into three separate packages (stitch-design / stitch-build / stitch-utilities) so coding agents can call Stitch's MCP server to do interface design. It's a concrete case of a major vendor applying the "skill marketplace" model to its own product line, and it's been flagged by automated trending-watcher bots two days running.
  • Tech stack: Agent Skills open standard + Stitch MCP server + the npx plugins/npx skills CLIs
  • Getting started: Medium — you need the Stitch MCP server configured first, then install into your coding agent with npx plugins add google-labs-code/stitch-skills

ysunyang4/Nexus 405

GitHub · TypeScript · AGPL-3.0

  • What it is: A local MCP control plane that lets web-based AI assistants like ChatGPT and DeepSeek reach your computer through a browser extension — running file reads/writes, terminal commands, and Git operations across hundreds of tools, with scope and permissions you set yourself.
  • Why it matters: Most MCP clients on the market only serve coding agents installed locally (Claude Code, Cursor). Nexus flips that: it's built for people who want to keep using the web version of ChatGPT but still have it act on their machine. It logs every tool call to a write-ahead log so a dropped connection won't replay actions, and sensitive files (.env, .pem, id_rsa) are masked by default. The developer posted about it themselves on the LINUX DO community forum, and it went from zero to 405 stars in five days.
  • Tech stack: Tauri 2 desktop app + Fastify MCP server (Streamable HTTP) + SQLite WAL + browser extension
  • Getting started: Medium — you need to install the Nexus desktop app and then the browser extension before web AI output can reach your local tools

mastra-ai/softwarefactory-template 64 (launched today)

GitHub · TypeScript · unlicensed

  • What it is: The "Factory" template Mastra unveiled on Show HN today — connect your GitHub repo and it turns issues into plans, implementations, and pull requests for a human to review, all running on a Factory Server you host yourself, with optional Linear, Jira, and Slack integration.
  • Why it matters: It's the same issue-in, PR-out, human-in-the-middle pattern as yesterday's Tencent LoopForge, but Mastra packaged it as a template where auth, storage, and sandboxing can each be swapped independently — not locked to one vendor. And Factory itself is built on Mastra's own framework, making it a public case of the company eating its own dog food.
  • Tech stack: Mastra framework + PostgreSQL/pgvector + Docker sandbox, with swappable auth/storage/sandbox providers
  • Getting started: Medium — npm create factory scaffolds the project, but you still need to configure GitHub App permissions and a model provider key before running your first issue

Notable Releases

Pydantic AI v2.70.0

Release Notes

  • Key changes: Fixed two high-severity ReDoS bugs — one in the regex that strips ANSI escape sequences under strip_ansi (GHSA-m4cr-m4qc-9hh3), and one in the built-in jwt secret-redaction pattern (GHSA-5m5q-5qj4-gx9h) — both fixed in pydantic-ai-harness 0.70.0. Also added StreamGuardrail (screens streamed output before the caller sees it), PruneToolResults, and forwarded service_tier to xAI.
  • Breaking changes: No hard break at the function-signature level, but the release frames this as "Compatibility Notes" and lists 14 behavior changes, including Groq's output-format detection and Anthropic's retry logic.
  • Impact: If your pipeline runs strip_ansi on model output, or uses the built-in redact_secrets/secret_data() jwt pattern on text an attacker can influence, either regex could previously be fed a few hundred KB of crafted input to stall the event loop for seconds to minutes before you upgrade to pydantic-ai-harness 0.70.0 — worth upgrading directly rather than weighing the risk first.

CrewAI 1.15.42

Release Notes

  • Key changes: Made ChromaDB an optional extra instead of a required dependency, and switched Knowledge's default backend to a local Qdrant instead of locking it to Chroma. Also raised the soupsieve version floor to 4.0, fixing CVE-2026-126190.
  • Breaking changes: If your code assumed Knowledge always uses ChromaDB — say, by calling the underlying Chroma client directly — the default backend is now Qdrant, and you'll need to explicitly specify which vector store to use.
  • Impact: Using CrewAI's Knowledge API as-is won't break, but if your deployment pulled in ChromaDB as a side effect of installing crewAI — say, for some unrelated purpose — you now need to add the extra explicitly to still get it installed.

Today's Takeaway

I used to assume agent-framework releases were mostly about adding features, but today Pydantic AI and CrewAI both shipped updates fixing holes introduced by what they pull in — one in a regex they wrote themselves, one in a third-party package's own CVE. The frameworks' own code wasn't the problem; the risk came up through the chain of dependencies they carry. It's a reminder that assessing an agent framework's risk surface means looking past its own commit log to the string of packages it drags along.

References