Illustration of a magnifying glass examining connected software blocks on a dashboard, representing how the AI Breakpoint Report methodology reviews and classifies AI vulnerabilities.

AI Breakpoint Report: Methodology

How SPB’s AI Breakpoint Report decides what counts: the CVSS 8.0 threshold, nine categories of AI software, counting rules, how data is collected and classified, and the count’s known limits.

The AI Breakpoint Report counts published vulnerabilities rated CVSS 8.0 or higher in software whose main purpose is AI. Each monthly edition covers one calendar month; the latest is the October 2026 edition. This methodology explains what qualifies, how flaws are collected and classified, and where the count has limits. It is versioned: when a rule changes, the change is logged at the bottom of this page.

Severity Threshold: CVSS 8.0 and Above

We count flaws scored 8.0 or higher on the Common Vulnerability Scoring System (CVSS). CVSS labels everything from 7.0 to 8.9 “High” and sets no official break at 8.0. The 8.0 line is SPB’s editorial choice, made to isolate the flaws attackers can most easily exploit at scale.

In practice, the line tends to separate two kinds of flaws. Those just below it more often need local access or cause one kind of damage, such as data theft or a service outage. Those at 8.0 and above are generally reachable over a network and can cause broad damage, up to full system compromise. In our 2026 data, about 62% of counted flaws can be exploited over a network with no login.

Flaws scored 7.0 to 7.9 still matter. We don’t collect or report them.

Which score we use. When a record carries a CVSS v4.0 score, we use it. Otherwise we use the v3.x score. Where NIST’s National Vulnerability Database hasn’t scored a flaw, we use the score from the CVE Numbering Authority that issued it. We don’t re-score flaws ourselves.

What Counts as AI Software

A flaw counts only if the affected product’s main documented purpose falls into one of nine categories.

AI Breakpoint · Methodology

The nine categories of AI software we count

Monthly editions use the plain-language name; the registry label is shown beneath it. The bar shows each category’s share of the 1,507 severe flaws counted from January through October 2026.

CategoryWhat it coversExamplesShare of 2026 flaws
Agent platformsRegistry: Orchestrate agents Platforms and frameworks that build and run AI agents Langflow, Flowise, PraisonAI 41%
MCP connectorsRegistry: Agent tool access Model Context Protocol servers and connectors that give agents tools and data MCP servers, Google MCP Toolbox 13%
Model-serving softwareRegistry: Serve inference Software that hosts and serves models vLLM, SGLang, Ollama 8%
AI gatewaysRegistry: AI gateway/control Proxies and gateways that route, meter or control model traffic LiteLLM, GitLab AI Gateway 5%
Model managementRegistry: Manage models/AI workloads Tools that track, version and deploy models and AI jobs MLflow 4%
AI/ML build toolsRegistry: Build AI/ML systems Libraries and tools for building and training models Hugging Face Datasets, NLTK, MONAI 7%
Model operationsRegistry: Operate models Tools for running and monitoring models in production LM Studio, GPT-SoVITS, rembg 1%
RAG and vector databasesRegistry: RAG/vector infrastructure Retrieval pipelines and vector stores that ground models in data Crawl4AI, mem0 5%
AI apps and chatbotsRegistry: Generative AI applications Chatbots, coding assistants and AI plugins Mistral Vibe, WordPress AI chatbot plugins 16%

Shares are rounded and may not total 100%. Monthly editions group RAG and vector databases with model operations as “Other.” Category totals for each month are in the October 2026 edition. Source: SPB AI Breakpoint Data.

What We Exclude

  • Software that runs around AI but isn’t AI, such as operating systems, browsers, general-purpose databases and GPU drivers.
  • Non-AI parts of mixed products. In a product that does many things, only flaws in its AI or MCP component count. A product that calls itself “AI-powered” doesn’t qualify on that basis alone: it counts as AI software only if it can’t do its main job without a model.
  • Malicious packages. A package published to attack people is an attack, not a vulnerability.

Two Types of Flaw

Every counted flaw gets one of two labels:

Most flaws we count are AI-stack: about 96% so far in 2026. The split is one way we track whether AI software is maturing.

Counting Rules

  • One flaw, one count. When a CVE record and a GitHub advisory describe the same flaw, it counts once.
  • Bundles. When one advisory bundles several CVEs, each CVE counts and the bundle doesn’t. When two advisories describe the same flaw (the same code path, attack and fix), it counts once, even if each carries its own CVE number. An advisory GitHub withdraws no longer counts.
  • Exploited means listed in CISA’s Known Exploited Vulnerabilities catalog. We don’t count vendor or researcher reports of exploitation unless CISA lists the flaw.
  • Month means the month the advisory was first published to GitHub’s Advisory Database, not the month it was updated, scored, patched or first disclosed elsewhere.

How the Data Is Collected and Classified

  1. Collect. An automated job runs daily against the full GitHub Advisory Database. That includes GitHub-reviewed advisories and the unreviewed records GitHub imports from NIST’s National Vulnerability Database. We read the whole database rather than a pre-filtered feed, so the AI filter is entirely ours.
  2. Filter. The job keeps records scored 8.0 or higher. It re-checks the previous 14 days for records whose scores arrived late. It flags a record as a possible AI flaw if it matches a product in SPB’s registry of AI software, names a major AI brand, uses an AI term such as LLM, MCP, inference or RAG, or is tagged CWE-1427, the weakness category for prompt injection.
  3. Enrich. The job removes duplicates and adds exploitation signals: CISA KEV status and the EPSS score, which estimates how likely exploitation is.
  4. Classify. Each flagged record goes to an AI model, Anthropic’s Claude, with a fixed set of written rules. The model labels it Counted, Not AI or Needs ruling, and records the product, maker, category and flaw type. When the model isn’t confident, it answers Needs ruling instead of guessing.
  5. Rule. An SPB editor decides every Needs ruling case and, before each edition, reviews borderline records by hand. When we rule that a product is or isn’t AI software, the ruling goes into the registry, so the same product is treated the same way every time after that.
  6. Log. Corrections are recorded, not erased. A record removed from the count, such as a duplicate, stays in our dataset with a dated note.

Disclosure: SPB uses Anthropic’s Claude in this workflow. Anthropic products also appear in the data, and they are counted under the same rules as every other maker’s.

Reporting Window and Late Records

Each edition covers one calendar month and publishes about two weeks after the month ends. A vulnerability’s first public record is often incomplete: severity scores can be added or revised days or weeks later, advisories get reviewed and corrected, and patches ship after disclosure. Waiting lets most of a month’s records settle before we count them.

Each edition states its data cutoff date. Records that arrive or change after the cutoff are added to the month they belong to in our year-to-date totals, so a month’s figure may rise slightly over time. Published editions aren’t rewritten; corrections are noted on the edition they affect.

The October 2026 edition used data through Nov. 18, 2026.

Brand 25

We track 25 leading AI companies every month, whether or not they had a counted flaw, so a zero is a recorded result rather than a gap. A zero doesn’t necessarily mean clean code. Some companies don’t publish CVEs for their AI products, and we check for that before reporting a zero. See the current Brand 25 rankings in the October 2026 edition.

Known Limitations

  • The filter can miss flaws. A record with no AI wording and no match in our registry never reaches classification.
  • We inherit published scores. We don’t re-score flaws, so a mis-scored record is counted as scored.
  • Nothing below 8.0 is collected or reported.
  • Mixed products take judgment. Deciding which part of a product counts as AI is an editorial call, and we log those calls in the registry.
  • Counts reflect disclosure, not only risk. A vendor that publishes many CVEs will look worse than one that publishes none. Rising counts partly reflect more researchers looking at AI software.

Corrections and Version History

If you think we got something wrong, such as a miscategorized product, a missed flaw or a bad number, email contact@securitypointbreak.com. We fix errors and note the change in the affected edition.

AI Breakpoint Report methodology version history
VersionDateChange
1.0Current First public version, published with the October 2026 edition
Total
0
Shares
Previous Article
Isometric illustration of an AI software stack, with connected blocks and a padlocked core module, representing the AI Breakpoint Report’s count of AI vulnerabilities.

AI Breakpoint Report: A Monthly Count of Severe Vulnerabilities in AI Software

Related Posts