airbag limits what a coding agent can do to your machine. A way for the agent to get past that limit is the report that matters most.
Report it privately through GitHub: Report a vulnerability (GitHub's private vulnerability reporting, on the Security tab). Please do not open a public issue, pull request or discussion about it, or post it elsewhere, until a fix is out.
A report helps most when it has:
- the airbag version (
airbag version) and the output ofairbag doctor; - the OS, the kernel and the agent, with their versions;
- the policy files in use and the
airbag runflags; - what the agent runs, and what you then see on the host;
- which promise in the README or the docs it breaks.
One problem per report. Reproduce it yourself first, and if a tool helped you find it or write it up, say so. A report we cannot reproduce, or one about code that does not exist, is closed.
airbag has one maintainer, so these are the times we can keep:
- an acknowledgement within 7 days;
- an update at least every 14 days until the report is fixed or closed;
- coordinated disclosure within 90 days of the report. If a fix needs longer, we agree on the date with you.
If 7 days pass with no answer, open a public issue that asks us to look at the advisory, with no details in it.
We fix it in private, publish a GitHub security advisory that names the fixed version, request a CVE through GitHub when the problem is in scope, and credit you unless you would rather not. airbag is a command first, though some of its Go packages can be embedded: either way, watch the repository's releases and security advisories to hear about fixes.
airbag is at 0.x. Fixes go into the latest 0.x release; older releases get
none. Until the first release is tagged, that is main.
The agent doing what airbag says it cannot. With a default or documented
configuration, an agent under airbag run that can:
- change your files or
$HOMEbeforeairbag apply, other than the agent state the docs say passes through, or run code on the host before you approve it; - read what airbag hides: credential files and variables, host sockets, a bound credential's real value, or airbag's own memory and environment;
- reach a host the allowlist and policy do not allow, including this machine, link-local or a cloud metadata address through an allowed name, or, after a secret read, reach what the read cuts off;
- read a watched secret file without the session being labelled
secret; - loosen the policy it runs under;
- get a
git pushpast the outbox, or makeairbag applywrite somethingairbag reviewdid not show; - push input into the terminal you return to;
- make a syscall, or open a socket family or netlink protocol, that the docs say the seccomp filter refuses;
- crash airbag's process on the host, or hold more there than the bounds the docs state, through the proxy, the forwards, the mirror or the control socket.
A way to change a release that reaches users (the binaries, SHA256SUMS,
install.sh or the release workflow) is in scope too.
So is a public Go package that fails open when it is used as
docs/composition.md says: the proxy reaching a host that
its allowlist, address guard or Gate refuse, or handing a bound credential's
real value to the client; the policy engine allowing what its rules deny; the
outbox or githubpr publishing a request that was not approved and claimed,
or publishing it twice. What the embedding host owns is not: its isolation
(whether traffic can get past the proxy), the labels it collects and reports
through Gate, and how durable its Recorder is.
The limits airbag documents are not vulnerabilities, unless they also break a promise the docs make elsewhere:
- airbag is not a VM. The kernel is shared, so a kernel bug the sandbox can reach is a kernel bug: report it upstream. A call the docs say the filter refuses that gets through is in scope, above.
- What the agent reads is sent to the model API. Its own file tools are not filtered, so a secret it reads can reach its model.
- What the policy allows: an allowed host used to send data out, a bound token
used on its hosts,
--allow,--pass-env,--nix-daemon. - The rest of the threat model: the proxy decides by the name the client asks for and does not see inside TLS, so a broad allowlist entry and domain fronting are ways out; bandwidth and the rate of new connections are not limited.
- The gaps the README and docs/ name: the two of the default mode that
--strictcloses, and the one it leaves when airbag runs as root; a socketcore_pattern; adefer:command called by its full path. - Command models are predictions. A script they read as
opaqueis not a bypass; the sandbox, the proxy and FUSE are the boundary for it. - On the macOS prototype, the limits docs/macos.md lists.
- Prompt injection of the agent itself.
- Hardening ideas with no way past the boundary shown. Open a public issue for those.