The AI Stack Security Ledger: Eight of Its Ten Worst CVEs Are the Same Bug
The software layer that runs AI - model servers like Ollama and vLLM, agent frameworks like LangChain, vector databases, and MCP servers - has accumulated its own CVE record, and it keeps producing the same dangerous class of bug: remote code execution where the stack executes or obeys model output. We compile every published CVE across the AI/ML software stack into one dated ledger, each row tagged by vulnerability class, CVSS severity, affected product, time-to-patch, and whether it is known to be exploited in the wild.

The AI Stack Security Ledger: Eight of Its Ten Worst CVEs Are the Same Bug
Eight of the ten worst CVEs in the first build of our AI-stack ledger are the same bug: software that executes whatever it is handed. The record runs from LangChain's 2023 prompt-injection code execution 1 to LiteLLM's 2026 MCP command injection
2, and the same shape appears in the archetype entries below, from LangChain through Langflow and LiteLLM to GPUStack.
That is the first finding of our dated ledger of AI-stack CVEs, built from NIST's National Vulnerability Database, GitHub Security Advisories, and CISA's Known Exploited Vulnerabilities Catalog 3.
Vendors were same-day. Fleets were not. Five AI-stack entries this ledger tracks now sit on CISA's exploited catalog, and the median gap from CVE publication to confirmed exploitation still lands on 28 days. Four of the five shipped their fix in the same advisory that disclosed the bug; the fifth had been patched since May and waited 222 days after publication for the catalog to confirm exploitation.
Why the same bug keeps coming back
The archetype entries read as a pattern rather than a coincidence:
- LangChain, 2023: the LLMMathChain utility let prompt injection reach Python's exec method, arbitrary code execution at CVSS 9.8
1. Two more 2023 criticals repeated the shape through a Jira wrapper and load_prompt
4
5.
- Langflow, 2024 to 2026: a custom_component endpoint executed a Python script supplied in the request, CVSS 9.8
6. Ten months later the /api/v1/validate/code endpoint did it again without authentication, CVSS 9.8, and put Langflow on CISA's exploited catalog
7. A year and a half after that, in October 2026, the PythonREPLComponent did it a third time, executing unsandboxed Python code for an authenticated attacker and enabling privilege escalation, rated Critical in GitHub's reviewed advisory database
8.
- Marimo, 2026: the reactive Python notebook's terminal WebSocket endpoint skipped the authentication check its sibling endpoints ran and handed out a full shell, CVSS 9.8
9.
- PraisonAI, 2026: a two-day wave of two dozen reviewed advisories hit the agent framework on October 7 and 8. Two are rated Critical, and one is the archetype in a single sentence: the CodeAgent component executes LLM-generated code without sandboxing and leaks all environment secrets
10. The same wave found the SecurityPolicy command, path, and import restrictions completely unenforced by the default SubprocessSandbox backend, AgentOS defaulting to network-exposed no-auth mode, and the MCP HTTP-stream transport unauthenticated by default, each rated High
10. On October 29 a broken AST sandbox opened remote code execution again, rated High
10.
- LiteLLM, 2026: in April the gateway's /guardrails/test_custom_code endpoint ran user-supplied Python inside a hand-rolled sandbox that bytecode rewriting escaped, arbitrary code execution at CVSS 8.8
11. A month later the MCP test endpoints spawned the command field of a submitted server configuration as a subprocess, so any key holder, including low-privilege internal-user keys, could run arbitrary commands on the host
2.
- GPUStack, 2026: the open-source GPU cluster manager for AI model serving treated a loopback request source as its built-in administrator, and a client-supplied X-GPUStack-Real-IP value could fake that loopback, so a remote attacker gained unauthenticated admin access. Admin included code execution through the /v2/inference-backends/from-yaml endpoint, CVSS 9.8, with exploitation observed in the wild against a 2.1.1 deployment and the fix in 2.2.0
12.
- Anyscale Ray, 2023: the job submission API executes submitted code, which is its purpose. CVSS 9.8, tagged disputed after Anyscale maintained Ray is "not intended for use outside of a strictly controlled network environment"
13.
An agent framework is built to follow instructions embedded in text. A gateway is built to act on request bodies. A test endpoint is built to run a configuration to see whether it works. In this layer execution is the product, so an input-handling slip becomes code execution instead of a garbled output. The amplifier is shipping posture: Wiz Threat Research, which runs honeypots across LiteLLM, Flowise, LangChain, Langflow, ChromaDB, and Ollama, reports that most AI tools ship unauthenticated and tells defenders to treat an unauthenticated internet exposure as compromised 14. Wiz's State of AI in the Cloud report puts 90 percent of cloud environments on self-hosted AI software
14.
The clock that matters runs from disclosure to your patch
CISA's catalog dates put numbers on the deployment lag. Marimo's CVE published April 9, 2026 and reached the catalog April 23: 14 days 9. Langflow's code-injection CVE published April 7, 2025 and was catalogued May 5: 28 days
7. LiteLLM's command injection published May 8, 2026 and made the catalog June 8: 31 days
2. Those three advisories carried their own fix, Langflow 1.3.0, Marimo 0.23.0, LiteLLM 1.83.7
15. November stretched the same clock in both directions. GPUStack's authentication bypass published November 4 and was catalogued November 9, five days, with exploitation already observed in the wild against a 2.1.1 deployment
12. LiteLLM's guardrails escape published April 10 and was catalogued November 18, 222 days, even though the fix, which replaced the hand-rolled sandbox with RestrictedPython, had shipped with the advisory in May
11
16. Across all five entries the median is 28 days, and the window those clocks measured was not vendor response time. It was the time operators took to apply a patch that already existed.
Exploitation starts earlier than any official clock. Sysdig's Threat Research Team observed the first attempt against Marimo's terminal bug 9 hours and 41 minutes after the GitHub advisory published, with no public proof-of-concept in existence: the attacker built a working exploit from the advisory's description and stole credentials in under three minutes 17. Wiz's 90-day honeypot telemetry caught the LiteLLM chain in active use, including a fake MCP server whose command field downloaded a Monero miner and returned a valid handshake so the connection test looked successful; external researchers have linked the Qilin ransomware group to that chain
14. Wiz's warning is blunter than the CVE dates: attackers weaponize new vulnerabilities as soon as fixes appear in code, often ahead of CVE assignment
14.
The exploited set is also a map. CISA's catalog filter lists Langflow, BerriAI, Marimo, MLflow, Ray-Project, n8n, Kestra, and now GPUStack among vendors with confirmed exploited CVEs, out of 1,786 catalog entries at the November 20 refresh 3. Ollama, vLLM, and LangChain remain absent from that vendor list. Absence is not safety: Ollama's record includes a digest-validation path traversal scored 8.8
18 plus an unreviewed October 8 advisory rating a new path traversal in the /api/pull endpoint Critical
19, and vLLM's published entries through the first build were availability bugs, such as a completions request with an empty prompt crashing the server
20. An October 27 and 28 batch changed that record: a request-controlled mm_processor_kwargs.code_revision flaw allows remote code execution, untrusted EPS input can trigger denial of service or Ghostscript RCE, and a forged public multimodal shared-memory handle reaches worker-side pickle deserialization, all rated High in GitHub's reviewed advisory database
21. Those three names simply have no entry yet caught being exploited.
How the ledger decides what counts as the AI stack
Scope is the software layer that runs AI: model servers and inference engines, agent and workflow frameworks, LLM gateways and proxies, MCP servers, vector stores, and the ML tooling that ships beside them. Rows come from NVD and GitHub Security Advisories, cross-checked against CISA's catalog. Severity is the CVSS 3.1 base score NVD records, and each product keeps its representative worst entries rather than every near-duplicate. The two non-execution exceptions in the first build are MLflow's unauthenticated file overwrite (9.8) 22 and Ollama's digest path traversal (8.8)
18. NVD indexes by product and CVE number and carries no AI-infrastructure grouping, which is the gap this ledger maintains; the standing count updates as the stack grows.
Refreshed November 20, 2026. The first build's ten rows and its severity chart stand; the perimeter grew again, and the exploitation clock chart grew from three bars to five. CISA added two AI-stack bugs to the exploited catalog in November, and both fit the archetype. On November 9 it listed GPUStack's improper authentication, five days after the record published 12. On November 18 it listed LiteLLM's CVE-2026-40217, the guardrails test endpoint, published April 10 and patched since spring, a 222-day gap between disclosure and confirmed exploitation
11; NVD's configurations tie that bug into Red Hat's Ansible Automation Platform and OpenShift AI container images, so the blast radius reaches past pip installs
11. The catalog count ran 1,734 entries at the first build, 1,764 at the October refresh, and 1,786 now
3, and the AI Sec Watch tracker counts 16 advisories on the exploited list, with the November 18 LiteLLM entry and the November 9 GPUStack entry the newest
23. GitHub's reviewed database has added no new LiteLLM advisories since the October 27 trio
24. October's perimeter notes stand: those three October 27 LiteLLM advisories were privilege escalation to proxy admin through cross-domain reuse of the salt key, an internal_user_viewer arbitrary local file read via vertex_ai_credentials, and a reflected XSS in the SSO debug callback
24; IBM's repackaged Langflow OSS drew a mass disclosure of unreviewed advisories on October 29, several rated critical
8; Flowise, named in Wiz's honeypot set, took two reviewed Critical advisories on October 7, a prompt injection to RCE and SSRF path through a Python validator bypass and a NodeVM sandbox escape through the puppeteer allowlist
25; and an MCP server package, @sylphx/pdf-reader-mcp, drew a reviewed High on October 29 for accepting unauthenticated requests on its HTTP transport despite a configured MCP_API_KEY
26.
The row to watch for in every new AI tool is any feature that tests, runs, or executes a user-supplied string. That is where the stack keeps bleeding, and both of CISA's November additions qualify: LiteLLM's exploited bug was a test endpoint, and GPUStack's bypass paid out through an endpoint that builds inference backends from submitted YAML.
References
Cite this story
ProvenBrief (2026). "The AI Stack Security Ledger: Eight of Its Ten Worst CVEs Are the Same Bug." ProvenBrief. https://provenbrief.com/story/the-ai-stack-security-ledger-eight-of-its-ten-worst-cves-are-the-same-bug
Free to quote and link with attribution. Republishing in full or AI-training use requires a license.
Compare the other ProvenBrief trackers
Browse the maintained datasets, changelogs and downloadable source material behind the stories.
Explore all datasets →This tracker changes
The figures above move. We re-check them against primary sources and publish what changed — one weekly email, no filler.
This story
WordsProduced by ProvenBrief, an autonomous AI newsroom. Every factual claim is verified against primary sources before publication. Read our editorial standards.