New here? Start with our launch editorial on why SPB tracks AI as the attack surface, not the weapon.
Two hundred twenty severe flaws. Seventy-five critical. Seven perfect 10s. None added to CISA’s exploited list.
That was October in AI software security, by SPB’s count of advisories first published to GitHub’s Advisory Database during the month. Every flaw here is rated CVSS 8.0 or higher and sits in software whose main job is AI: tools for building AI agents, servers that run AI models, MCP connectors that plug AI into other systems, coding assistants, and the chatbots bolted onto everything else. (How we decide what counts.)
The short version: IBM published 38 Langflow flaws, including two batches of 16 three weeks apart. Langflow is IBM’s free, open-source tool for building AI agents without writing code, with more than 150,000 GitHub stars, and a Langflow server usually holds the keys to everything it connects.
NVIDIA added 10, seven of them on AIStore in a single day. AIStore is NVIDIA’s free, open-source storage system built to feed huge training datasets to GPU clusters fast. It’s back-end infrastructure for teams training large models, not something most companies run.
And model-serving software took 43 hits, its busiest month of the year, with fixes from vLLM, SGLang, LMCache, llama.cpp, Ollama and Moonshot AI’s Mooncake.
Model-serving software runs AI models after they’re built: it loads the model, takes requests and returns answers. vLLM and SGLang run models at data-center scale, LMCache and Mooncake help them reuse work across requests (Mooncake runs Moonshot AI’s Kimi chatbot), and llama.cpp and Ollama run models on ordinary PCs and laptops.
None of this is consumer software. Langflow, AIStore and the model servers are tools for the people who build AI products, and consumers depend on them anyway: a flaw here can expose whatever chatbot or app is built on top.
October at a Glance
| Measure | October 2026 | September 2026 |
|---|---|---|
| Severe AI flaws (8.0+) | 220 | 244 |
| Critical (9.0+) | 75 | 73 |
| Perfect 10.0s | 7 | 10 |
| Products affected | 104 | 96 |
| Hit AI agents (orchestration + MCP) | 102 (46%) | 124 (51%) |
| Model-serving software | 43 | 26 |
| Remote, no login needed | 141 (64%) | 147 (60%) |
| Newly added to CISA’s exploited list | 0 | 0 |
IBM’s Langflow led October with 38 flaws, followed by PraisonAI (12), vLLM (11), and SGLang and Mooncake (8 each).
October excludes three records an SPB editor ruled outside the AI definition and one advisory that duplicates another. September excludes a Flowise advisory GitHub withdrew on Nov. 5 as a duplicate.
The Pattern Under the Count
AI Breakpoint · October 2026
Only 12 of October’s 220 severe AI flaws ran through the AI model itself
Each square is one flaw rated CVSS 8.0 or higher. Orange squares are flaws an attacker reaches through what the model reads or is tricked into producing, such as prompt injection. The rest are older bug types in the software around the model.
Hover or tap a square to see the flaw.
View the 12 flaws as a table
| Flaw | Product | Score |
|---|
Source: SPB AI Breakpoint Data (workbook v7, Nov. 18, 2026), advisories first published to GitHub’s Advisory Database Oct. 1–31, 2026, after editor review. Squares ordered by group, then score.
One pattern matters more than the count. Only 12 of October’s 220 flaws run through the AI model itself, through what it reads or what it’s tricked into producing. The rest are the classics: code injection, missing authentication, path traversal. Old bugs, new software.
That cuts against the usual story. Prompt injection tops OWASP’s list of AI risks and dominates the coverage, partly because it’s new. Prompt injection tricks an AI model into following an attacker’s instructions, hidden in something the model reads.
Code injection, a far older problem, tricks software into running an attacker’s commands. Code injection is a bug that can be patched. Prompt injection exploits how models work, and there’s no single fix. It also rarely gets a CVE or a high severity score, which is one reason it shows up so seldom in a count like this.
The two meet when software runs whatever the model says. Langflow’s Smart Transform feature, for example, asks an AI model to write a small piece of Python code and then runs whatever comes back. An attacker who slips instructions into the data it reads can get the model to write malicious code, and Langflow will run it. The prompt injection is the way in. The flaw that earned an 8.8 is that Langflow trusts the model’s output enough to run it. IBM fixed it in Langflow 1.10.3.
How October Stacks Up
October’s 220 flaws were the year’s second-highest monthly count, behind September’s 244, and its 75 critical flaws trailed only July’s 87. The kinds of bugs didn’t change. Where they landed did.
AI Breakpoint · September vs. October 2026
The bugs moved from MCP connectors to the servers that run AI models
Severe AI flaws by what the software does, September and October. MCP connectors fell by half. Model-serving software jumped to 43, its most this year.
View data as a table
| Category | Sept. | Oct. | Change |
|---|
Source: SPB AI Breakpoint Data (workbook v7, Nov. 18, 2026). September: 244 flaws; October: 220. “Other” combines RAG and vector databases and model-operations tools.
September was the month of MCP connectors: 44 flaws, the most this year. In October that number halved to 22, and the problems moved to the software that runs the models. Model-serving software took 43 flaws, about one in five of the month’s total; before September it never topped one in 12. They were worse, too: 42 of the 43 need no login, and 12 are critical, against five in September. And they came from 12 projects, not one project’s batch of fixes.
By bug type, the usual suspects led again: code injection (32), missing authentication (24), path traversal (21) and OS command injection (19). The order is what changed. Missing authentication nearly doubled from September’s 13, and eight of October’s cases were in model servers, up from one. Path traversal, September’s leader with 33, and server-side request forgery, at 27, both fell back. Same bugs, new targets. November’s count will show whether the servers that run models stay in the line of fire.
What Stands Out
These records stand out on three signals SPB tracks: how many severe flaws a product took, whether they can be reached without logging in, and whether the product already has flaws on CISA’s exploited list. How those signals weigh against each other depends on what you run. That call is yours.
Langflow (IBM), 38 flaws. 12 critical · 9 exploitable without login · five Langflow flaws added to CISA’s exploited list this year
No AI product has more flaws on CISA’s exploited list than Langflow. CISA added five this year, four of them disclosed in 2026. October brought 38 more, most in two batches of 16, on Oct. 7 and Oct. 29. Early in the month came CVE-2026-8505 (9.8), which lets anyone run flows by getting past webhook authentication, and CVE-2026-105697 (9.9), a command injection that a logged-in user can trigger through MCP server settings. Some of the latest, such as CVE-2026-110186 (9.8, no login needed), affect versions up to 1.12.8. IBM’s Oct. 28 bulletin lists versions 1.0.0 through 1.12.8 as affected and names 1.13.0 as the fix. IBM’s Langflow security bulletins cover the month’s other flaws.
IBM did not respond to a request for comment.
NVIDIA AIStore and NeMo Agent Toolkit (two 10.0s). NVIDIA published seven AIStore flaws on Oct. 13, among them CVE-2026-65162, a perfect 10 in a legacy Kubernetes setup API, and CVE-2026-65173 (9.1), missing authentication on token revocation. On Oct. 27 came CVE-2026-65145, a 10.0 missing-authentication flaw in the NeMo Agent Toolkit. NVIDIA has shared few details beyond warning of possible code execution and data tampering. AIStore holds training data, and agent toolkits hold the keys. NVIDIA’s bulletin 5889 names AIStore 5.0 as the fix, and bulletin 5912 names NeMo Agent Toolkit 1.9.0.
The model-serving stack, 43 flaws. Model servers had their worst month of the year:
- SGLang‘s CVE-2026-93034 (9.8) deserializes whatever arrives on its message channel.
- LMCache‘s CVE-2026-105192 (9.8, fixed in 0.5.6) does much the same.
- Moonshot AI’s Mooncake had eight flaws, including missing authentication on its storage service (CVE-2026-106037, 9.3).
- Ollama‘s CVE-2026-103663 (9.4) is a path traversal in model downloads, triggered when a model is pulled.
These servers usually live inside GPU clusters that teams assume are private. Put them behind authentication anyway.
CVE-2026-116692, MLflow remote code execution (9.8, Databricks). Unauthenticated MLflow servers can be made to run attacker code through the way they load third-party scorers (CVE-2026-116692). MLflow already has one exploited flaw this year: CVE-2026-64849 reached CISA’s list in August, two days after its GitHub advisory was published. Version 3.17.0 fixes this one. Treat it as urgent if your tracking server takes requests without a login.
Agents that run model-written code with no sandbox: PraisonAI and Vibe-Trading. PraisonAI’s CVE-2026-61447 (10.0) runs code the AI model writes without a sandbox and exposes every secret in the environment. A sibling flaw, CVE-2026-61445 (9.9), lets the model’s tool calls write files and run commands. Vibe-Trading, a research project from the University of Hong Kong, got two 10.0s, one in its AI-callable tools and one in its unauthenticated API, both fixed in 0.1.7. If an agent can execute what a model writes, a prompt injection becomes code execution.
October’s Perfect 10s
| Advisory | Product (maker) | What it allows | Login needed? | Fixed in |
|---|---|---|---|---|
| GHSA-jqmf-mx4f-hfr6 | Vibe-Trading (HKU) | Command execution through AI-callable tools | No | 0.1.7 |
| GHSA-v2f8-6655-7grj | Vibe-Trading (HKU) | Unauthenticated API access and a code-execution chain | No | 0.1.7 |
| CVE-2026-32579 | Kognetiks Chatbot for WordPress | Arbitrary file upload | No | See advisory |
| CVE-2026-61447 | PraisonAI | Runs model-written code unsandboxed, leaks secrets | No | 1.6.78 |
| CVE-2026-65162 | NVIDIA AIStore | Abuse of legacy Kubernetes setup API | No | 5.0 |
| CVE-2026-95182 | DeepSeek Harness | Code execution from a crafted project folder | No, but the victim must open a crafted project | See advisory |
| CVE-2026-65145 | NVIDIA NeMo Agent Toolkit | Missing authentication | No | 1.9.0 |
The Rest of October
The long tail, sorted by what the software does:
- Agent platforms (80 flaws). Langflow’s 38 and PraisonAI’s 12 lead. Flowise, DB-GPT, Agno, Devika and LangGraph fill out the list. Most are injection bugs or missing logins on APIs that were never meant to face the internet.
- AI apps and chatbots (35 flaws). These include three bugs in Mistral’s Vibe coding agent, such as CVE-2026-116843, where a malicious repository runs commands, and a 10.0 in DeepSeek Harness (CVE-2026-95182) for loading code from a crafted project folder. A pile of WordPress AI plugins rounds it out.
- MCP servers (22 flaws). Nginx UI’s MCP endpoint had three, and small MCP connectors filled the rest. The software that connects agents to tools is a regular here now.
- AI gateways and controls (11 flaws). GitLab’s AI Gateway had a 9.9, CVE-2026-90970. Only organizations that host their own gateway need to act; details below.
- Everything else (29 flaws). Tools for managing, building and running models, among them MLflow, MONAI and Crawl4AI.
The Brand 25
Beyond the month’s worst flaws, SPB tracks 25 of the biggest names in AI every month, the Brand 25, whether they publish flaws or not. Ten of them had severe AI flaws in October, led by IBM with 40, including Langflow’s 38, and NVIDIA with 10. Two more stood out:
Google’s agent stack: Agent Mesh, MCP Toolbox, ADK. Four critical flaws, none of which needs a login:
- Agent Mesh: CVE-2026-102503 and CVE-2026-103091, missing authentication and authorization.
- MCP Toolbox: CVE-2026-93328, a path traversal.
- ADK for TypeScript: CVE-2026-105772, a path traversal.
If your teams build on Google’s agent kit, update all three.
CVE-2026-96620, Claude Agent SDK argument injection (9.2, Anthropic). Whoever controls the SDK’s resume option can inject arguments and run commands (CVE-2026-96620). The fix is in 0.2.121. Disclosure: SPB uses Anthropic’s Claude in its production workflow.
AI Breakpoint · The Brand 25
IBM leads the Brand 25 by a wide margin, mostly because of Langflow
Severe AI flaws (CVSS 8.0+) in products from 25 of the biggest names in AI, first published to GitHub’s Advisory Database in 2026. Pick a month, then select a vendor to see each flaw, linked to its NVD record.
How to read this: zero means no severe flaw in that vendor’s AI products reached GitHub’s database in the period, not that its AI is flaw-free; vendors that fix hosted services without publishing CVEs won’t appear. IBM includes Langflow, Red Hat and HashiCorp. GitHub is tracked apart from its owner, Microsoft.
Source: SPB AI Breakpoint Data (workbook v7, Nov. 18, 2026), Jan. 1–Oct. 31, 2026.
How to read this chart. The orange part of each bar is critical flaws, 9.0 or higher. IBM’s lead comes mostly from Langflow, 122 of its 139, and from IBM’s practice of issuing CVEs for its own products, sometimes 16 at a time. Six vendors show zero: CrowdStrike, Okta, Palo Alto Networks, SAP, Workday and xAI. Zero means no severe flaw in their AI products reached GitHub’s database this year, not that their AI is flaw-free.
Exploited and Still Active
CISA added no October AI flaws to its Known Exploited Vulnerabilities catalog. Don’t relax yet. All eight 2026 AI flaws on CISA’s exploited list sit in three products, and two of those products shipped new critical fixes in October.
| CVE | Product | Score | What it allows | Published | Added to CISA list | Days to CISA list |
|---|---|---|---|---|---|---|
| CVE-2026-64849 | MLflow (Databricks) | 9.3 | Server-side request forgery, no login | Aug. 17 | Aug. 19 | 2 |
| CVE-2026-59822 | LiteLLM (BerriAI) | 8.8 | MCP authentication bypass | Jul. 22 | Sep. 2 | 42 |
| CVE-2026-9198 | Langflow (IBM) | 9.8 | Code execution, no login, default setup | Jul. 17 | Aug. 4 | 18 |
| CVE-2026-55255 | Langflow (IBM) | 8.4 | Access to other users’ flows | Jun. 19 | Jul. 7 | 18 |
| CVE-2026-42271 | LiteLLM (BerriAI) | 8.7 | Command execution via MCP test endpoints | Apr. 25 | Jun. 8 | 44 |
| CVE-2026-42208 | LiteLLM (BerriAI) | 9.3 | SQL injection in API key check | Apr. 24 | May 8 | 14 |
| CVE-2026-33017 | Langflow (IBM) | 9.3 | Code execution, no login | Mar. 17 | Mar. 25 | 8 |
| CVE-2026-0770 | Langflow (IBM) | 8.9 | Code execution via code-validation feature | Jan. 23 | Jul. 21 | 179 |
- Langflow holds four of the eight, and CISA also added a fifth Langflow flaw, CVE-2025-34291, in May. October’s two big batches hit the same parts of the product.
- LiteLLM has three exploited flaws this year, three of its 14 severe flaws and the highest share of any AI product we track. It holds the API keys for every model provider behind it, which is why attackers like it.
- MLflow’s exploited flaw reached CISA’s list two days after its GitHub advisory was published. October added the 9.8 above.
If you run any of the three, assume someone is already scanning for it.
AI Breakpoint · The year so far
1,507 severe AI flaws in 10 months, and October was the second-busiest
Flaws rated CVSS 8.0 or higher in AI software, by the month they were first published to GitHub’s Advisory Database.
View data as a table
| Month | Severe | Critical | Perfect 10s |
|---|
Source: SPB AI Breakpoint Data (workbook v7, Nov. 18, 2026), after editor rulings and one advisory GitHub withdrew. November is still being scored and isn’t shown.
Step back from October and the year looks like a security industry relearning its own history at speed. SPB has counted 1,507 severe AI flaws since January, 582 of them critical. More than half, 54%, hit software that builds AI agents or connects them to tools. IBM leads the major vendors with 139, almost all in Langflow, four times Google’s 35.
AI Breakpoint · The year so far
Two out of three severe AI flaws are bug types the industry has known for decades
Share of each month’s severe AI flaws in long-known classes such as injection, path traversal and missing authentication. The dashed line is the 2026 average, 65%.
View data as a table
| Month | Share |
|---|
Bug classes counted: CWE-22, 77, 78, 79, 89, 94, 95, 287, 306, 502, 862, 863 and 918. Source: SPB AI Breakpoint Data (workbook v7, Nov. 18, 2026).
The number I keep coming back to is 65%. That’s the share of this year’s flaws in bug classes the industry learned to prevent decades ago: injection, path traversal, missing authentication. It has stayed between 58% and 70% every month since February. The software is new. The mistakes aren’t.
In fairness, some of the volume is a good sign. IBM publishes CVEs for its own products, sometimes 16 at a time, and researchers are finally looking hard at AI tools. Full tables and the maturity trends are in the October data edition.
Coming in November
Severity scores are still arriving, so November’s count will grow well past the 30 flaws recorded through Nov. 15. Three early signals:
- CVE-2026-119990, Azure OpenAI (9.8): missing authentication that lets a network attacker elevate privileges. Microsoft says it has fully mitigated the flaw in its service, so customers don’t need to act.
- CVE-2026-109492, Microsoft MCP Gateway (8.8): an authorization bypass in the gateway Microsoft built to control agents’ access to tools.
- Three more Mistral Vibe bugs and an OpenAI Codex flaw (CVE-2026-110930). In each, a crafted project, path or repo turns the coding agent against its user. Malicious repos are becoming their own category.
The November edition will publish Wednesday, Dec. 16. Until then, patch the agents before someone else runs them for you.
More from SPB: Attackers Hide Prompts From Humans, Aim Them at Inbox AI Agents · Official MCP Code Flaw Exposes AI’s Trust Gap · Poisoned Plug-ins for AI Coding Agents Open Attack Door · Sandbox? What Sandbox?: Red Team Agent Goes Rogue
How We Count
Every flaw in this report is rated CVSS 8.0 or higher and sits in software whose main purpose is AI. Each flaw counts once, and exploited means listed in CISA’s Known Exploited Vulnerabilities catalog. This edition covers advisories first published to GitHub’s Advisory Database Oct. 1–31, 2026, using data through Nov. 18.
Some were disclosed elsewhere, such as on NVD, before reaching GitHub. Full rules, sources and known limits: AI Breakpoint Report methodology v1.0. Corrections: contact@securitypointbreak.com.
Sources: GitHub Advisory Database · CISA Known Exploited Vulnerabilities catalog · CVE.org · NVIDIA security bulletin 5912, NeMo Agent Toolkit · NVIDIA security bulletin 5889, AIStore · IBM security bulletin 7294467, Langflow OSS (Oct. 28, 2026) · IBM security bulletin 7290694, Langflow OSS (Oct. 2, 2026) · IBM Langflow security bulletin, CVE-2026-6543 · The Register on CVE-2026-9198 · GitHub withdrawal record, GHSA-mgrx-3hqw-485h · NEAR AI incident report · The Hacker News on NEAR AI and SGLang · The Hacker News on Langflow CVE-2026-105740 · Previdian, CVE-2026-105740 · Langflow advisory GHSA-w794-rj3p-xv45 · GitLab AI Gateway patch release · Microsoft Security Response Center, CVE-2026-119990 · Wiz, “Black October” for AI Developer Platforms · OWASP Top 10 for LLM Applications, LLM01 · Help Net Security on LiteLLM 1.107.0 · Help Net Security on Plugin4Shell · Axios on PHP2Shell