Scanning the 302 most popular MCP servers, skills and plugins
138 MCP servers, 64 skills and 100 plugins, ranked by downloads or GitHub stars, each pinned to an exact version or commit and scanned with Bawwab's static rules. Scan date: 2026-11-20.
Headline counts
144 flagged by static rules ยท 0 confirmed harmful after review. The rule verdicts below are what the scanner said before anyone read the code; each target's page shows the reviewed outcome next to it.
Every target that was not clean was reviewed before publication. Reviews were AI-assisted for 105 of the 144 flagged targets (Claude reading the pinned code); 25 of them were also checked by a maintainer. Each finding names its reviewer. Review outcome for the 144 flagged targets:
- Reviewed: expected behaviour: 87
- Reviewed: false alarm: 57
How this was done
- Selection. MCP servers are ranked by weekly npm or PyPI downloads (official modelcontextprotocol/servers entries always included); skills and plugins by GitHub stars of the repository that hosts them, with at most two per repository after Anthropic's official ones. Full method.
- Pinned. Each target is scanned at an exact package version or git commit, listed on its page.
- Download only. Packages are fetched with Bawwab's own fetchers and read as files. Nothing is installed or run.
- Static rules only. The verdict comes from Bawwab's pattern and code-flow rules, without the AI review step Bawwab normally adds, and without running anything in a sandbox.
- Reviewed. Each flagged finding carries a reviewer's status: confirmed, known false alarm, or expected for what the tool does.
Categories
Each finding is tagged with the public threat categories its rule maps to: the OWASP Top 10 for LLM Applications, the OWASP Top 10 for Agentic Applications, MITRE ATLAS techniques and CWE weaknesses. The tags describe the kind of risk the rule looks for, not a confirmed attack. Rules that only say part of a target was not checked are left untagged on purpose. Sources:
- OWASP Top 10 for LLM Applications, version 2025, retrieved 2026-11-20
- OWASP Top 10 for Agentic Applications, version 2026 (published 2025-12-09), retrieved 2026-11-20
- MITRE ATLAS, version 2026.10, retrieved 2026-11-20
- MITRE CWE, version 4.21, retrieved 2026-11-20
Limits
- Static rules see code patterns, not intent. "Not malicious, needs a look" means a pattern worth reading, not an accusation.
- No AI review was used, so findings that Bawwab's judge would normally clear are shown as raw rule hits with a reviewer's note instead.
- Very large files are skipped by the scanner and reported as such; code that is downloaded or generated at run time is not seen.
- A clean result means no rule fired at this version. It is not a guarantee, and later versions may differ.
- Popularity numbers are a snapshot and can be inflated (bots, CI, star campaigns); they decide what was scanned, not how it scored.
Reproduce
git checkout 5b813032b16847c0b0fed41186443d7bd30fb483 report/run_scan.sh # reads the pinned report/targets.json, writes report/out/results.json uv run python report/build_site.py
Disputes
If you maintain one of these packages and think a result is wrong, open a dispute (each target page has a link with the target filled in). Say which finding, why it is wrong, and link to the code. We re-check and update the page and the review note.