AI security, week of 9 to 15 November 2026

This is the seven days to 15 November 2026, and the thread was the server behind the agent. The tools we stand up to give a model hands, MCP servers, RAG back ends, model gateways and the app builders around them, kept trusting whoever reached them: an absolute path read any file, a spoofed header reset a rate limit, a request body swapped a gateway’s own credentials, and a connection marked read only still wrote. The anatomy walks the sharpest case, an unauthenticated route in a ComfyUI extension that copies an attacker’s file over a custom node and a second route that restarts the process to run it. The 9 November post covered the window before this one so nothing repeats here, and where the ARTEX framework returns it is a new FBI advisory, not the campaign I wrote up earlier. As the week closed Amazon put a frame around all of it: its decoy sensors recorded attackers’ own agents combing the internet for exactly these exposed tools, hunting the API keys inside them.

At a glance

By layer

Inference and serving

Agent frameworks and MCP

AI applications and UIs

MLOps and data pipelines

Also patched: tsParticles MCP server 4.3.1 and 4.3.2 (CVE-2026-63655, CVE-2026-76892, CVE-2026-76893, denial of service in HTTP mode), Dagu 2.11.0 (CVE-2026-102943, MCP path-containment bypass), ezBookkeeping 1.6.0 (CVE-2026-86041, the same spoofable X-Forwarded-For allowlist), pgEdge RAG Server and Go LLM library (CVE-2026-125325 to CVE-2026-125328, provider key fragments in error messages and unauthenticated corpus reads), and NLTK 3.10.4 (CVE-2026-126303, unbounded CPU in a grammar parser).

Case files

Cursor: a stolen model-provider key reached the extension marketplace (14 November)

Cursor is an AI code editor built on VS Code, and Fireworks AI is a model inference and training provider it used. The credential the attackers started from was an access key stored at Fireworks.

With it they copied a read-only bucket that Cursor used to share data with Fireworks, holding older coding-agent traces, mostly from 2025, from users without Privacy Mode on. Some of those traces still held unredacted API credentials, which opened more of Cursor’s systems, including one that could publish to it’s extension marketplace. On 29 October the attackers placed malicious copies of Anthropic’s Claude Code extension (versions 2.1.312 and 2.1.313) on the cache that serves the marketplace, where they sat about eleven hours, and the build tried to plant an SSH key and switch on remote access.

Per BleepingComputer, Cursor revoked the Fireworks key, rotated the exposed credentials, rebuilt the marketplace, moved inference off Fireworks and removed the builds. It estimates up to 5,000 users may have received a malicious build, fewer than 0.05% had traces exposed and fewer than 0.01% had credentials taken, and says production systems, official releases and Privacy Mode users were untouched. Researchers reported coding errors that likely kept the backdoor from working where SSH was not already on. Cursor’s own incident report is the primary account.

OWASP LLM Top 10 2025: LLM03 Supply Chain. MITRE ATLAS: AML.T0010 AI Supply Chain Compromise, with AML.T0055 Unsecured Credentials for the keys left in the traces.

Jet Admin: a prompt injection read the platform’s .env (13 November)

Jet Admin is a hosted platform for building internal tools and dashboards, and the attacker in this one needs only a free account.

CVE-2026-123965 describes a prompt injection that let that account reach the platform’s .env configuration files and pull the credentials out of them, which NVD scores 9.9. The vendor record says it affects all versions and that the fix is server side, a sandbox component added around 29 October, so there is no release for a self-hosted reader to install. The CVE and a coordinated advisory (Italy’s CSIRT tracking VA-26-317-01) were published on 13 November.

For anyone running an AI feature that can read configuration or secrets, the control is to keep those secrets out of the model’s reach, because untrusted input that reaches the model is input that can ask for the .env. It is the same gap the week’s credential-hunting agents were counting on.

OWASP LLM Top 10 2025: LLM01 Prompt Injection and LLM02 Sensitive Information Disclosure. MITRE ATLAS: AML.T0051 LLM Prompt Injection.

Anatomy of the week: a ComfyUI extension that copied a file into itself

ComfyUI is a node-based interface for image models, and ComfyUI-Easy-Use is one of the most installed custom-node packs for it. CVE-2026-123790 was coordinated through CERT/CC (VU#944667) by Nguyễn Huy Hoàng and made public on 12 November, with a tagged fix, v1.4.2, following on 14 November; the CVE record carries no score and the finder self-rates it 9.8. A default install listens only on loopback, but the --listen flag and most hosted GPU images make it reachable, and that reach is the only real precondition.

  1. The attacker uploads a file of Python code. ComfyUI’s own POST /upload/image takes no authentication and writes the bytes under input/ with no check on type or content.
  2. They call POST /easyuse/save/{name} with {name} set to custom_nodes%2F<node>%2F__init__.py. The handler splits on the first slash and the percent-encoded slash survives routing, so the folder type it reads is custom_nodes and the name is an installed node’s entry point.
  3. The handler checks that the request’s subfolder stays under a base directory, then joins the body’s filename on afterwards. An absolute or ../ filename makes os.path.join discard the base, so it points back at the uploaded payload, which becomes the copy source.
  4. get_full_path("custom_nodes", "<node>/__init__.py") resolves only because that file already exists, so the write is an overwrite of the node’s __init__.py, taking its extension from the source.
  5. shutil.copyfile copies the payload over the entry point.
  6. GET /easyuse/reboot, also unauthenticated, calls os.execv and restarts the process on demand. On restart ComfyUI imports its custom nodes, the replaced __init__.py runs, and the attacker has code execution as the ComfyUI process, with no user interaction.

Where it could have stopped

The shape to look for is a path or permission check that runs before the dangerous value is fully built, on a service written as though it would only ever hear from localhost. In your own systems, list every unauthenticated route that writes a file or restarts a process, and confirm the check lands on the final resolved path.

From the research side

Checklist for Monday

References

By layer:

Case files:

Anatomy:

Research:

Notes on sources: