Daily brief
Cursor Traces Its Extension Breach to a Vendor Key, JetBrains Puts Coding Agents in TeamCity, and Browser Use Raises Its Default Price
By Steven Sorensen · 16 Nov 2026

At a glance
- Cursor says the fake Claude Code extension in its marketplace began with an access key a vendor exposed, and it has told every affected customer directly.
- JetBrains has an experimental plugin that installs Codex or Claude Code inside TeamCity builds and hands them provider keys only while the build runs.
- Browser Use now runs a far more expensive model when you pass no model of your own.
- No new stable release from Claude Code, Copilot CLI, Codex or Gemini CLI, and both Cursor changelogs are unchanged.
Cursor posted its update on Nov 13, older than our usual window, but it is the first full account since the Oct 29 notice. JetBrains describes the TeamCity plugin on its blog, and Browser Use tagged v1.0.6 on GitHub at about 7:51 PM ET Sunday. The GitHub Changelog has nothing new since Nov 13 in the afternoon.
Treat today as a hygiene and cost check day: confirm nobody on your team still carries the fake extension or its SSH key, decide whether coding agents belong in your CI builds before someone wires them in, and pin the Browser Use model if your scripts rely on the default.
Top Stories
1. Cursor traces the fake Claude Code extension to a key a vendor exposed
Cursor's Nov 13 update explains how a malicious copy of Anthropic's Claude Code extension reached its VS Code extension marketplace on Oct 29. An attacker obtained an access key that Cursor had issued to Fireworks, and Fireworks exposed it. With that key the attacker copied a read only storage bucket of coding agent traces, mostly from 2025 and only from sessions not in Privacy Mode. A small number of those traces held credentials that had not been redacted, and Cursor says less than 0.01% of customers had API credentials exposed.
From there the attacker reached some SaaS tools Cursor uses, limited support data and some credentials users had sent to support. On Oct 29 a stolen credential let the attacker place tampered copies of the extension, v2.1.312 and v2.1.313, on the cache that serves the Cursor marketplace. They were up for about 11 hours, and Cursor says at most 5,000 users who requested those cached versions may have received one. Anthropic did not publish the copy, and Open VSX, the marketplace source, never had the malware.
Cursor says the app, its updates, other extensions and customer repositories were not affected, and no release signing keys were exposed. It moved all inference off Fireworks, rotated credentials and rebuilt the marketplace, and the last attacker activity it saw, on Nov 1, was blocked.
Practical dev impact: If you installed or updated the Claude Code extension from the Cursor marketplace on Oct 29, follow the Oct 29 security notice rather than only uninstalling, because Cursor says uninstalling alone may not be enough. Everyone else should keep secrets out of agent sessions that are not in Privacy Mode.
2. JetBrains lets TeamCity builds run Codex and Claude Code
JetBrains describes an experimental AI Agents plugin for TeamCity on its TeamCity blog. When a build starts, the plugin checks whether the selected agent is on the build agent and installs it if it is missing, so you call codex or claude from a normal Command Line build step. Provider keys live in a TeamCity connection, and the plugin gives them to the build agent only for the length of the build and removes them afterward. It supports OpenAI, Anthropic, OpenRouter, LiteLLM and JetBrains Air.
Each run also leaves a record. The plugin publishes prompt, completion, cached and total token counts as TeamCity statistics, and it saves the agent trajectory as HTML for reading and JSON for analysis, in hidden build artifacts under .teamcity/ai-agents/. JetBrains warns that these artifacts may contain prompts, source code and model output, so access to them should be restricted like any other sensitive build artifact.
Practical dev impact: This makes agent tasks such as code review or failed build analysis an ordinary build step on shared agents. Start in one project, restrict who can open the trajectory files, and chart the token statistics from day one so a cost spike shows up early. The plugin is experimental, so expect changes.
3. Browser Use v1.0.6 makes its pricier model the default
Browser Use v1.0.6 changes what Agent(task) and ChatBrowserUse() run when you pass no model. The GitHub release says they now use bu-4.0 instead of bu-fast. The notes say plainly that this costs more, because BU 4.0 tokens cost 15 times as much as bu-fast and run at high reasoning instead of medium in the default mode.
The notes give the way back. Pass llm='bu-fast' to keep the previous model and price, llm='bu-4.0-mini' for simple tasks at the lowest cost, or llm='bu-4.0-pro' for the hardest tasks.
Practical dev impact: If you have scheduled jobs or tests that build an agent without naming a model, your bill changes the day you upgrade. Pass the model explicitly before you move to v1.0.6. When we checked at 6:03 AM ET Mon Nov 16, PyPI still listed v1.0.5 as the latest build, so the GitHub release has not reached a plain install yet.
Practical Impact Analysis
Our read: the Cursor update has the widest reach because so many developers use the Cursor marketplace, even though only a small group was exposed. It is also the only story with a cleanup step that may still be open, so it leads despite its Nov 13 date. The other two are smaller, but both put agents where they run without a person watching.
The common thread is that agent records and defaults now carry real risk, because Cursor's incident started with traces that held credentials, TeamCity trajectories may hold source code, and Browser Use changes cost through a default you may never have set. Codex posted v0.171.0 alpha builds over the weekend whose notes list only the version, so leave them for a test machine and stay on the current stable release.
Tutorial
- Open a test shell. Every command below only reads files, versions or public package data.
- Step 1 is the macOS and Linux form of the check in Cursor's Oct 29 notice. On Windows, use the PowerShell lines from that notice.
- Step 2 compares your Browser Use build with what PyPI lists now, and step 3 finds agents that may pass no model.
- Step 4 shows whether the agent CLIs a TeamCity build would call are on this machine.
# We have not run this script end to end on every platform. It prints and checks, and changes no files.
# Run it in a test shell. Nothing is written to your shell profile.
# 1. Look for the attacker key named in Cursor's Oct 29 notice (reads one file if it exists).
if [ -f "$HOME/.ssh/authorized_keys" ]; then
if grep -l svc-telemetry "$HOME/.ssh/authorized_keys"; then
echo "svc-telemetry key found, follow the Cursor security notice"
else
echo "No svc-telemetry key found"
fi
else
echo "no authorized_keys file"
fi
# 2. Snapshot at 6:03 AM ET Mon Nov 16 2026:
# GitHub release v1.0.6, PyPI latest still v1.0.5. No install is pinned.
if python3 -m pip show browser-use >/dev/null 2>&1; then
python3 -m pip show browser-use
else
echo "browser-use not installed in this Python"
fi
python3 -c 'import json, urllib.request; d = json.load(urllib.request.urlopen("https://pypi.org/pypi/browser-use/json")); print("PyPI latest", d["info"]["version"])'
# 3. Python files that create an agent or the Browser Use chat model. Check each one for an llm argument.
grep -l -r --include="*.py" -e "Agent(" -e "ChatBrowserUse(" .
# 4. Agent CLIs on this machine.
if command -v codex >/dev/null; then
codex --version
else
echo "no codex on PATH"
fi
if command -v claude >/dev/null; then
claude --version
else
echo "no claude on PATH"
fi
echo "Done. Nothing was installed or changed on disk."
If PyPI shows a newer version than your snapshot, read its notes before you upgrade.
Sources:
- An update on our recent security incident (Cursor)
- Security notice on the fake Claude Code extension (Cursor)
- Run AI Agents in TeamCity (JetBrains)
- Browser Use v1.0.6 release
- Browser Use on PyPI
- Codex releases
- Cursor changelog
- Cursor CLI changelog
- GitHub Changelog
Recommended AI prompt
Copy this paragraph into ChatGPT, Claude, Gemini, Grok, or whatever AI you use.
You are my developer tooling lead for Mon Nov 16, 2026. Story 1: Cursor's Nov 13 update says a key it issued to the vendor Fireworks was exposed, which led to tampered copies of the Claude Code extension, v2.1.312 and v2.1.313, on the Cursor marketplace cache for about 11 hours on Oct 29, and its notice says uninstalling alone may not be enough. Story 2: JetBrains describes an experimental TeamCity plugin that installs Codex or Claude Code in builds, gives provider keys only for the length of a build, and saves token counts and agent trajectories as artifacts. Story 3: Browser Use v1.0.6, tagged on GitHub, makes the more expensive BU 4.0 model the default when no model is passed, while PyPI still listed v1.0.5 at this check. Ask which editors, CI system and agent libraries my team uses, then give a short checklist for each story, and link the official posts instead of inventing settings, flags or prices.