AI security review — ecc
Plugin: affaan-m/ECC · 250k stars · MIT
Reviewer: Fable 5.1 · Date: 6 September 2026
Grade: A · Risk: none
What it is
ECC (by Affaan Mustafa) is the largest of the three: a harness-native plugin for engineering teams. In this catalogue it carries seventy skills, sixty-eight agents, fifty-six commands, three instruction files, and one MCP convention — around two hundred files, plus optional lifecycle hooks you can turn off in the plugin’s config.
What I looked at
I read the manifest and all bundled components. The manifest is standard: MIT, an empty mcpServers block, and a userConfig that lets you disable the hooks or pick a profile. Nothing runs at install time.
Most of the bundled agents are code reviewers — a Java reviewer, a PHP reviewer, a TypeScript reviewer, and so on. This matters for reading the scan, because a code-review agent’s text is full of the very patterns a scanner is trained to catch.
Findings
Every flag resolved to a false positive once read in context.
The reviewers list things like Runtime.exec(), shell_exec(), and “command injection” — but as items on a checklist of what to look for in someone else’s code, not as commands the agent runs. The agent-evaluator was flagged on a line that actually describes a read-only Bash allowlist (grep, cat, ls) it restricts itself to. The build-resolver agents run rm -rf against build caches (.next, .gradle/build-cache) — routine cleanup. A few media and data skills call documented APIs (fal.ai, ElevenLabs, X), which is what they are for. Two skills tell you to add a key to ~/.claude.json or settings.json; that is setup instruction, not a change made behind your back.
| Pattern flagged | Where | Verdict |
|---|---|---|
exec / shell_exec / “command injection” | code-review agents | review checklist, not executed |
rm -rf .next / .gradle | build-resolver agents | build-cache cleanup |
fetch / requests to APIs | media & data skills | the tool’s stated purpose |
“add to settings.json” | config docs | setup instruction |
A full-text pass for piped installers, obfuscation, credential reads, and exfiltration came back clean.
Verdict
A large but orderly engineering toolkit. Its size and its many code-review agents make the raw scanner noisy; read in context, none of it is dangerous. The hooks are optional and disclosed in the config. Grade A.
_Produced by agentmods.dev — automated AI security review. Reviewer model: Fable 5.1. Date: 6 September 2026. How this review works: https://agentmods.dev/about#security _