| Capability | RepoSentry | Dependency & SAST scanners | Manual code review | Just trusting stars & vibes |
|---|---|---|---|---|
| Checks dependencies against known CVEs | nobody does this by hand | |||
| Reads the repo's actual code for malicious intent, not just its manifest | pattern rules, no judgement | |||
| Flags content aimed at manipulating an AI coding agent (prompt injection, hidden Unicode, untrusted MCP configs) | not built for this threat | reviewers aren't looking for it | ||
| Filters placeholder/test-fixture matches out before they count as real findings | usually raw regex hits | |||
| Public due-diligence context (repo age, maintainer history, open issues) | only if you go look | |||
| Plain-English report a non-developer can act on | raw output, built for engineers | depends who's reviewing | ||
| OWASP Top 10 + CWE mapping for a compliance workpaper | CVE/CVSS only, usually | |||
| Time to a result | needs setup/CI first | hours, if it happens at all |
These describe how each category of tool is generally built, not a claim about any specific product's exact feature set today - your dependency scanner may do more than the row above credits it for.
What's actually new here
Checks for attacks aimed at your AI agent, not just you
Increasingly, the code you clone gets read by an AI coding assistant before a human ever does. That's a threat surface most scanners were never built for - prompt-injection phrasing addressed to an "AI"/"assistant", hidden or invisible Unicode text, and MCP server configs pointing at untrusted remote endpoints. RepoSentry checks for it explicitly, in its own report section.
Every high-severity finding is checked against the real file
A regex match, or an AI's own read of a file, is a claim - not a fact - until it's checked. Every AI-sourced finding, and every critical/high/medium secrets match, gets re-opened and confirmed against the actual cloned code before it can affect your score. A downgrade is logged with its reason, never a silent retraction - see /security.
One report, two audiences
The same scan produces a plain-English verdict a non-technical founder can act on, and an OWASP/CWE-tagged, CVSS-aligned finding table a security or compliance reviewer can drop straight into a workpaper. Most tools pick one audience.
Paste a GitHub URL, get a real answer in minutes
No install, no CI integration, no waiting on someone else's calendar.
Scan a repoSee /methodology for exactly what a scan checks, and /security for why RepoSentry itself is safe to run.