Supply chain security at agent speed.
Coding agents add packages faster than anyone can review them. Bitbison analyzes each package before it is installed, and the agent acts on the result: a fixed version, a different package or the one it asked for.
Reuses what you have
agentnpm install moment
bitbisonYour organization already uses date-fns in 14 projects.
agentUsing date-fns 3.6.0 instead.
Granted
Uses a fixed version
agentpip install requests==2.31.0
bitbison2.31.0 has an advisory, fixed in 2.32.0.
agentTaking 2.32.3.
Granted
Replaces a malicious package
agentcargo add time-sync
bitbisonBitbison has blocked time-sync 0.2.4: it uploads developers’ .env files. Your organization already uses chrono.
agentUsing chrono instead.
Blocked
Set it up once.
Then do nothing.
Success means you deploy Bitbison and do nothing else. Your agents check, choose and explain every package on their own.
Start with a plugin
Install the Bitbison plugin in your coding agent to begin.
No code is sent to us
Bitbison sees which packages your agents ask for and why. Your code stays where it is.
No access to your code
Bitbison does not ask for access to your repositories or your source.
Your team, once
- Install the Bitbison plugin in your coding agents.
- Run
bitbison setupso package managers download through your organization’s registry proxy. - Adopt each project’s existing lockfiles.
Your agents, every time
- Check what your organization already uses before adding something new.
- Take a fixed version, or mitigate, when an advisory is found for a package.
- Choose another release when Bitbison blocks one.
- Record why they took each package.
You control the policies that decide what is granted, what is blocked and what waits for a person’s approval.
Static and dynamic.
Detected at the source.
Bitbison reads every package and runs it, then connects what the package did to the code that did it. An attack is found where it starts.
Not just the source
Bitbison reads each package’s code, then builds and runs it.
Not just the build
Bitbison records every program, file and connection and traces each one back to its cause.
Not just a baseline
Bitbison judges a package by what it does, the first time it sees it. It needs no history to compare against.
xz-utils 5.6.0. The backdoor was not in the code anyone read. It was hidden in test files, decoded by the build scripts and linked into the library. Bitbison traces the library that shipped back to those test files. See the xz run
Read, built
and run first
Bitbison makes dynamic analysis possible at scale, built on our breakthrough in practical full systems provenance. It builds and runs each package on a clean, sealed machine and records every program it starts, every file it opens and every host it talks to.
solana-py-utils 1.98.2, published to PyPI on November 18, 2026, as Bitbison recorded its build: nine detections and the one host it should not have reached. See the whole run
A registry proxy
on every download
Your organization gets its own registry proxy. Package managers download through it, and it serves a release only when it is granted. The proxy covers downloads that agents did not request, and does not rely on agents following instructions.
An agent’s command
An install, an upgrade or an npx. A clean release and its dependencies are granted in the same step.
A developer or a script
The same rules hold for anyone on the machine, whatever started the download.
A dependency’s install script
Code a package fetches while it installs comes through the registry proxy too, and is refused without a grant.
Checked before
it is reported
A blocklist waits for someone to report a package. A waiting period hopes someone else notices first. Bitbison checks each package itself.
Bitbison finds compromises first
All published incidents- PyPI · agent-framework 1.26.1 · November 2026
A credential sweeper at every Python start.
The 1.26.1 wheel adds a .pth file that starts a credential sweeper at every Python start and uploads what it finds to a webhook.site URL absent from the project’s source.
Analyzed 6 minutes after 1.26.2 was published, 58 minutes before the first public advisory. Verdict: denyRead the review - npm · @wppconnect-team/wppconnect 2.4.1 · November 15, 2026
A hidden module that uploads secrets once a day.
Importing 2.4.1 runs a hidden “diagnostics” module that uploads environment variables, .env and config files and WhatsApp session tokens to a Telegram bot.
OSV credits Bitbison as the finder (MAL-2026-20044). Verdict: denyRead the review - npm · @ethersphere/bee-js 13.3.0 · November 9, 2026
An official release that installs a remote shell.
An official bee-js release pushed outside the project’s process runs code from an external server during installation. It sends the Swarm Desktop key and installs a persistent remote-shell agent.
Analyzed 5 minutes after it was published, 71 minutes before the first public advisory. Verdict: denyRead the review
Found only by Bitbison
No advisory in OSV as of November 18, 2026- Terraform · skybound1/skybound 2.6.2
A provider that opens a reverse shell.
The skybound1/skybound provider opens a reverse shell at every Terraform start.
Verdict: denyRead the review - PyPI · dynamic-mcp 1.0.9
A package that opens a remote shell to the internet.
dynamic-mcp opens a public Cloudflare tunnel to an unauthenticated remote shell and sends the URL to the publisher.
Verdict: denyRead the review - cargo · log-upgrades 0.1.2
A crate that adds an SSH key to the machine.
log-upgrades uploads .env and other files it sweeps up, and appends an SSH key to authorized_keys.
Verdict: denyRead the review - Release · xz-utils 5.6.0 · 2024
A backdoor hidden in test files.
The xz-utils 5.6.0 release tarball shipped with a covertly injected backdoor. It was hidden in test files, decoded by the build scripts and linked into the library.
A known attack from 2024. Bitbison detects it from the build alone. See what it did
Every decision,
with its reason
For each package your agents add, the console records who asked, which agent, which project and why. You don’t have to watch your agents. The record is there when you want it.
Know what is installed,
where and by which agent
Bitbison records every development package your organization installs: who asked for it, through which agent, on which machine and for which project.
Works with your agents
and package managers
A plugin for each coding agent, and a registry proxy your package managers already know how to talk to.
Coding agents
- Claude CodePlugin
- CodexPlugin
- opencodePlugin
- PiPlugin
Package ecosystems
- npmnpm, npx
- PyPIpip, uv
- crates.iocargo
- Go modulesgo
- RubyGemsgem, Bundler
- NuGetdotnet
- Maven CentralMaven, Gradle
- AndroidGradle
- OCamlopam
- RCRAN
- TerraformProviders and modules
- Claude Code plugins
New ecosystem support is added every week.
Sign up
Early access to Bitbison Code. Choose your region, then confirm your email.
- Sign up
- Confirm your email
- Join the waiting list
Check your email
We sent a link to . Open it and press Confirm my email to send your request. The link works for two days.
If the email hasn’t arrived, check your spam folder or send a new link. A new link replaces the previous one.
Questions
Do I have to review anything?
No. Your agents check, choose and record a reason for each package on their own. A person is asked only to adopt a project’s existing lockfiles the first time, approve an audit or override a block.
Does Bitbison see our code?
No. No code is sent to Bitbison, and Bitbison has no access to your repositories.
Does it slow my agents down?
A release Bitbison has already checked is granted in the same step the agent asks for it. A release nobody has asked for before is built first, and the agent carries on when the build finishes.
What happens when a package is blocked?
The agent is told why the package is blocked and what resolves the block: a fixed version, another release, a package your organization already uses or proof of work for a mitigation. It picks one and keeps working.
What data do you capture?
Your dependencies: which packages are asked for, who asked, for which project and why. And your audits, which are scoped to your organization. They are the proof of work when a vulnerability is mitigated, and they put known security risks in your agents’ context. All of it is kept in the region you choose when you sign up: America or Europe.
