AI agent stack vulnerabilities: what the public record holds
472 known vulnerabilities across 179 packages that AI agents are built from: model SDKs, agent frameworks, and MCP servers and SDKs. Counted from public OSV records as of November 13, 2026.
Published by Drel · Method · Limits · Download the data (CSV) · Records as of November 13, 2026
A team that builds an agent takes on the flaws of everything the agent is built from: the model SDK, the agent framework, and every MCP server it is given. This page counts what the public advisory record says about those packages. It is a count of published records, made by code. Nobody at Drel found, tested or reviewed the individual vulnerabilities.
Findings
- As of November 13, 2026, public advisory records describe 472 known vulnerabilities in 179 AI SDK, agent framework and MCP packages on npm, PyPI and Go.
- 292 of the 472 (62%) are rated critical or high by their own advisory: 78 critical and 214 high.
- 325 of them were published in 2026 so far, against 99 in all of 2025.
- The first vulnerability in an MCP-named package was published on May 29, 2025. In the third quarter of 2026 alone, 94 were. In total 267 vulnerabilities affect 145 MCP-named packages.
- For 76 of the 472 (16%), the record names no version that fixes the flaw.
- Separately, the record holds 174 reports of malicious code published under a package name, 170 of them under a name that says MCP. These are not counted as vulnerabilities above.
By publication date
Each vulnerability is dated by its earliest public advisory.
| Year published | All | MCP-named packages | SDKs and frameworks |
|---|---|---|---|
| 2023 | 19 | 0 | 19 |
| 2024 | 29 | 0 | 29 |
| 2025 | 99 | 39 | 60 |
| 2026 (to November 13, 2026) | 325 | 228 | 97 |
By severity
Severity is the advisory's own: its label where it has one, otherwise the level of its CVSS 3 score. Where several records describe one vulnerability, the highest level counts. A record with neither is “not given”; no level is guessed.
| Severity | Vulnerabilities | Share |
|---|---|---|
| Critical | 78 | 17% |
| High | 214 | 45% |
| Medium | 144 | 31% |
| Low | 29 | 6% |
| Not given in the record | 7 | 1% |
Packages with the most records
A long list is not a verdict on a package. Widely used projects with an active disclosure process collect more advisories than projects nobody examines, and a package with none may simply never have been looked at.
| Package | Registry | Vulnerabilities |
|---|---|---|
| litellm | PyPI | 45 |
| transformers | PyPI | 32 |
| mcp-atlassian | PyPI | 30 |
| langchain | PyPI | 23 |
| llama-index | PyPI | 22 |
| pydantic-ai | PyPI | 14 |
| pydantic-ai-slim | PyPI | 14 |
| n8n-mcp | npm | 11 |
| fastmcp | PyPI | 10 |
| llama-index-core | PyPI | 10 |
| mcp | PyPI | 10 |
| @modelcontextprotocol/sdk | npm | 9 |
| langchain-core | PyPI | 9 |
| @aborruso/ckan-mcp-server | npm | 7 |
| langchain-community | PyPI | 7 |
By registry: 300 on PyPI, 152 on npm, 22 on Go. A vulnerability that affects packages on two registries is counted once in each.
Method
- Source. The public OSV records for npm, PyPI and Go, copied on November 13, 2026. OSV aggregates the GitHub Advisory Database, the PyPA advisory database, the Go vulnerability database and the OpenSSF malicious-packages reports.
- Packages. The model SDKs, agent SDKs and agent frameworks on Drel's lists, plus every package whose name says MCP (the
@modelcontextprotocol/scope, or “mcp” as a separate word in the name). - One vulnerability, one count. Records that share an identifier or an alias are one vulnerability. The same flaw often appears as a GitHub record and a PyPA record under one CVE.
- Left out. Withdrawn records. Malicious-package reports are counted separately, never as vulnerabilities.
- Fixed version. A vulnerability has one when any of its records names a fixed version for at least one affected package.
- Who counted. Code, from the records as published. No person reviewed the individual records, and Drel tested none of these packages.
Every row is in the CSV, with its identifiers, packages, severity and a link to the OSV record, so any figure here can be recounted.
Limits
- These are published records, not all flaws. A flaw nobody reported, or one fixed without an advisory, is not here.
- The name rule misses MCP servers whose package name does not say MCP, and includes any package that uses the word for something else.
- SDKs and frameworks that are not on Drel's lists are not counted, and registries other than npm, PyPI and Go are not read.
- “No fixed version named” describes the record. A maintainer may have shipped a fix that the record does not mention.
- More records in a year can mean more flaws, more packages, or more people looking. The data does not say which.
Cite this page
Drel, “AI agent stack vulnerabilities: what the public record holds,” records as of November 13, 2026, https://drel.ai/ai-agent-stack-vulnerabilities. The figures change when the advisory copy is refreshed; the address does not. Quote the date with the number.
The underlying records belong to their sources and are used under their licenses: CC BY 4.0 for the GitHub, PyPA and Go advisory databases, and Apache 2.0 for the OpenSSF malicious-packages reports.
Frequently asked questions
- How many known vulnerabilities are there in MCP servers?
- As of November 13, 2026, public OSV records describe 267 known vulnerabilities in 145 packages whose name says MCP, on npm, PyPI and Go. MCP servers whose package name does not say MCP are not in that count.
- Are these vulnerabilities that Drel found?
- No. Every record comes from a public advisory database through OSV. Drel counted the records with code and did not find, test or review the individual vulnerabilities.
- Does a package with many advisories mean it is unsafe to use?
- Not by itself. Widely used projects with an active disclosure process collect more advisories. What matters for a given system is whether the version it locks is affected and what the agent can do through that package.
- How often do the numbers change?
- Whenever the advisory copy behind this page is refreshed. The page states the date of the records it counted; quote that date with any figure.
Review the agent, not only its dependencies
An AI security review covers what the agent can reach and do, the attack paths through its tools, the controls in place and the decision your AI Committee signs.
A note on scope: Drel reviews assessed systems against documented architecture, configuration and intent. It does not ingest live telemetry from production environments. Dispositions reflect the assessed system at the time of review and the re-assessment triggers that govern when the disposition must be revisited.