Insygna's free Agent Report Card: score agents pre-prod

Insygna launched a free AI agent security score on August 10: connect an agent's repository, get back a 0-100 rating across six dimensions plus findings down to the file and line. We looked at what it actually measures, and the useful takeaway is not the tool. It's that five of the six dimensions are boring build hygiene you can check yourself today.

What actually happened

Per the company's announcement, the Agent Report Card is a free public service that scans an agent repo and scores it 0-100 across six areas:

1. Secret Exposure — credentials in the repo or image
2. Dependency Vulnerabilities — known CVEs in what you pulled in
3. Code Security — the usual static analysis
4. Container Hardening — root user, exposed ports, capabilities
5. LLM Security — measured against the OWASP LLM Top 10
6. Image Security — the base image you built on

You get a findings list, version history, and a shareable "Insygna Verified" badge. It feeds a paid platform that assigns agents identities across Teams, Slack, Copilot, Claude and ChatGPT — so read the free tier as a funnel, which is fine as long as you know it.

Insygna also published a number worth quoting carefully: the median agent it has tested so far scores below 50 out of 100, and roughly six in ten fall below 50. That's the vendor's own sample of self-selected submissions, not a market survey. Directionally believable; not a statistic to put in a board deck.

Why an AI agent security score matters for your business

Count the dimensions again. Exactly one of six is agent-specific. The other five are the same supply-chain checks that applied to every container you shipped before anyone said "agent." If your agent repo scores badly, the fix is usually npm audit, a non-root USER line, and getting an API key out of a committed .env — not an AI strategy.

A repo scan is not a runtime test. This looks at code at rest. It does not watch what your agent does with a live tool call, and it cannot tell you that an approval gate is reachable around. That's a different control layer — the kind we described when AWS open-sourced Dogwood. Use both; don't mistake one for the other.

Steal the rubric, skip the badge. Six named dimensions with a pass/fail per agent is a perfectly good internal gate. Wire it into CI, make it block the merge, and you've built the same control without shipping your repository to a vendor you met on a Tuesday. Which is the actual question here: to score your agent, something has to read your agent.

A verification badge is marketing until someone re-runs it. Self-attested security scores age badly — one new dependency and the number is stale. If you display one, pin the date.

Sources: Insygna announcement via AIThority, OWASP Top 10 for LLM Applications.