Changelog: LLM Guard — archived · promptfoo — ownership changed · 7 of 21 vendors now acquired →

The Ledger That Scores
Coverage. Not Claims.

LLM Threat scores 21 LLM-security vendors and 10 open-source tools against the OWASP LLM Top 10 (2025) and MITRE ATLAS. Each profile marks all 10 risks as Covered, Partial, Not covered, or Unverified, and every one of those cells was checked and dated by hand. Last verified 2026-07-13.

See every vendor Open the coverage matrix

coverage_matrixlive

OWASP LLM Top 10 →

Covered Partial None Unverified 21 vendors →

17/21
Cover LLM01 Prompt Injection: 2/21
Cover LLM05 Improper Output Handling: 7/21
Vendors acquired: 210
Cells scored by hand: 23,165★
Stars on promptfoo, top tool

Why this exists

Every Vendor Site Sounds The Same. This Grid Shows What They Actually Cover.

Most vendor sites just reprint marketing bullets. We map every vendor onto a fixed grid — ten OWASP risks, four plain verdicts — and name the exact feature behind each score.

"Not covered" and "Unverified" aren't the same verdict. Not covered means we checked and it isn't there. Unverified means the vendor claims it but no source confirms it — a question for the demo call, not a win.

Market coverage by OWASP LLM risk — as of 2026-07-13
OWASP LLM Risk	Vendors covering it
LLM01 Prompt Injection	17/21
LLM02 Sensitive Information Disclosure	16/21
LLM03 Supply Chain	6/21
LLM04 Data and Model Poisoning	5/21
LLM05 Improper Output Handling	2/21
LLM06 Excessive Agency	8/21
LLM07 System Prompt Leakage	5/21
LLM08 Vector and Embedding Weaknesses	2/21
LLM09 Misinformation	8/21
LLM10 Unbounded Consumption	4/21

Of the ten risks, prompt injection has the best coverage — 17 of the 21 vendors we track handle it — while improper output handling is nearly untouched, with 2 covering it, per our July 2026 review of public vendor documentation.

Six pillars

Six Ways Into One Grid

AI Red Teaming

Tools and services that attack your model before a real adversary does, hunting injection paths, jailbreaks, and data leaks.

5 vendors · 4 tools

Guardrails / LLM Firewall

Request-time filters that sit between users and models, blocking injected prompts, unsafe output, and PII leaks before they land.

7 vendors · 5 tools

AI-SPM / Runtime Protection

Posture management with live detection: who's running which model where, and whether anything is attacking it right now.

5 vendors · 0 tools

AI Governance / Model Risk

Inventory, policy mapping, evidence, and audit workflows for GRC and model-risk teams — governance rather than guardrails.

1 vendors · 0 tools

Agentic AI Security

Controls for AI that acts — tool-calling agents and MCP traffic — where excessive agency turns a bad prompt into a bad action.

3 vendors · 1 tools

Open-Source Tools

Public GitHub projects for testing and defending LLMs, tracked with live stars, license, and repo status so a dead project can't sneak into your stack.

10 tools

Broadest coverage

Who Covers The Most

Ranked by OWASP LLM risks scored Covered. Each chip is a vendor's coverage fingerprint — ten cells, one per risk.

01 Lakera Guardrails / LLM Firewall · 8 covered · 2 partial 02 Noma Security AI-SPM / Runtime Protection · 5 covered · 4 partial 03 Straiker Agentic AI Security · 5 covered · 5 partial

No vendor pays for a better score — independence is the whole product.

Every cell is scored by hand against OWASP and MITRE ATLAS.

Each funding and repo figure carries a date and a source.

Sources: OWASP LLM Top 10 (2025), MITRE ATLAS, vendor documentation, GitHub API. See Data & Sources. Last verified 2026-07-13.