State of Agent Security · Q3 2026

We scanned 19,132 agent tools.
30% have a high or critical severity finding.

Not "could be risky." A concrete finding: a leaked secret, an unsafe exec path, an injected instruction in a tool description, a lethal-trifecta flow, with a file and a line, in a tool an agent is being told to connect to right now. That is 5,738 of 19,132 scanned tools (live figure), each counted once. 694 of them (3.6%) carry at least one critical finding.

Why this is the number that matters

The agent-security conversation in 2026 has been about identity (who is this agent?) and authorization (what is it allowed to do?). Both are being solved. But neither says whether the tool itself is safe to run. A perfectly-authenticated agent connecting to a perfectly-authorized MCP server can still hand it a secret, or call a tool whose description quietly tells the model to exfiltrate. 30% is the size of that blind spot, the third axis, the one nobody else is measuring. So we did.

What "flagged" means

At least one high or critical finding

We do not count low/medium noise, and findings that live only in a tool's test/example files are scored at a fraction. A tool is not judged on its test suite. A tool with both a critical and a high finding is counted once.

Only scanned tools count

The denominator is tools that received a verdict. Registry entries we could not fetch or scan (7,871 MCP entries alone) are listed in the Index as skipped and are in neither the numerator nor the denominator.

MCP servers: 34.7% flagged

4,281 of 12,351 scanned MCP servers carry a high or critical finding; 425 (3.4%) carry a critical one.

Every number recomputes

Each scan is a signed (Ed25519) verdict you can re-derive byte-for-byte offline against our public keys. Not a "trust us" survey; an attestation. Re-derive any figure here yourself.

The counter-example: it can be done right

The flip side of 30% is that a real top tier exists and is being cleared. At the end of Q3, 21 tools were AgentAvow Certified: verified build provenance, zero critical/high, no manifest drift, full offline-recomputable coverage, every check public. Among them were:

✓sigstore✓react✓axios✓undici✓vue✓@langchain/core

As of Q3 2026; the live Certified list has changed since. The gap between 30%-flagged and 21-Certified is the whole story: the ceiling is reachable. Most tools just are not near it yet. See the Certified gate →

If you build agent tools

Scan it (free, no account), see exactly what an agent inherits by connecting, fix what matters, and, if it is clean, put a signed badge on your README that anyone can verify. The 30% is not an indictment of developers; it is an unaddressed surface, and closing it is mostly low-hanging fruit.

Methodology

  • · Corpus: 19,132 tools with a verdict across GitHub repos, npm, PyPI, crates, container images, Hugging Face models, MCP servers, and skills, scanned by the public AgentAvow engine (12 detection categories at the time, 14 today; per-finding severity, exact file:line). Skipped and unfetchable entries are excluded from every figure.
  • · "Flagged" = at least one high or critical finding on shipped (non-test) code. Buckets are exclusive: critical, high-only, clean. flagged = critical + high-only, so no tool is counted twice.
  • · Every verdict is a signed, offline-recomputable attestation. Keys at agentgraph.co/.well-known/jwks.json; conformance vectors public.
  • · Live figure: the flagged-stat endpoint. This page renders it live, so it never drifts from the number it cites. Each count equals the total of the matching severity filter in the Index.

Figures update as the corpus grows. Last rendered from the live endpoint.