Skip to the content

Check what agents upload to paste sites and webhooks

Beta

Have AgentGarm read what launched agents send to paste sites, file sharing and webhooks, record or stop secrets, and fix tools that refuse its certificate.

Built, not yet claimed preview: this guide describes a package that has not passed its release gate.

List the paste sites, file-sharing sites and webhooks your agents should not quietly send secrets to, and AgentGarm reads each request a launched agent's tools send them, on this machine, before it leaves. At the end you can read what was found, stop uploads that carry a secret on the sites you choose, and get a tool that refuses AgentGarm's certificate working again.

1. Know what is read and what never is

Nothing is read until a policy lists sites, in security_boundaries.network.inspect. For those sites only, agentgarm-session opens the HTTPS connections that agents started with agentgarm run make, with a certificate from a web inspection CA that agentgarmd keeps on this machine, apart from the one it uses for model providers. It reads each request's address (path and query) and the body being uploaded, in memory, and scans them with the same detection as prompts: the credential formats and card numbers the prompt redaction guide lists, also inside form fields, multipart parts, compressed bodies (gzip, deflate, zstd) and base64. Then it forwards the request, records it or stops it. Request headers are not scanned.

Never read or kept: what the site sends back; any site no policy lists; model providers' APIs, subscription back ends such as claude.ai, identity providers' sign-in, update services and AgentGarm's own, which no policy can list; your browser and apps you run yourself; and the body, which is never stored, logged or uploaded. A finding keeps the site, the rule id (for example github_token), a count, where in the request it was (body, path, query, form, multipart, base64) and what was done, never the secret. The request's path is kept only with a finding or a refusal, on this machine only, with query values and any part holding a secret replaced by ….

Not covered in the beta: agents not started with agentgarm run (hooks only), MCP servers' own connections, the browsers a browser agent's shim starts, responses, and HTTP/3, which a launched agent cannot use, so its tools fall back to TCP.

How sure (every level is built, not yet claimed, until the package's release gate): on macOS a launched agent's sites in Block mode are B (controlled), since the sandbox sends every connection through the proxy and a tool that refuses the certificate is stopped; Audit and Alert are C (observed). On Linux it is C: Landlock limits connections by port, not address, so a tool can reach a site on the proxy's port directly, and before Linux 6.7 on any port. Windows is built and shown as C until its device run.

Beta. Outbound inspection is a public beta. It leaves the beta when, in the owner's own use: fewer than 1 % of connections to listed sites fall back to going uninspected; fewer than 1 false alarm per 1,000 inspected requests; the performance budgets hold (added latency p99 under 3 ms for a request with a body up to 64 KiB on a connection already open, under 15 ms for a task's first connection to a site, plus 2 ms per MiB of body held for a scan in Block, with agentgarmd and agentgarm-session together still idle under 1 % CPU and 80 MB of memory); and the owner's walk through these guides on a Mac. To tell us how it worked, the dashboard's Outbound inspection screen has Send feedback: it shows a report of counts and versions only, never a body, a path or a site outside the presets, and copies it for you. The site has no feedback form yet: send the report as the beta page's Feedback section says.

See also

Check that protection is on

See whether AgentGarm's services are up, what assurance level each agent session has and why, and what to do about a bypass alert.

2. Turn it on

On your own machine, agentgarm inspect enable --preset paste_sites --agent claude-code adds a preset to Claude Code's policy in Audit mode, the default (--mode takes audit, alert or block; without --agent it changes default.yaml). The presets are lists AgentGarm keeps and checks: paste_sites (15 sites, such as pastebin.com and paste.rs), file_sharing (14, such as file.io and transfer.sh) and webhooks (15, such as webhook.site, hooks.slack.com and discord.com). It prints the change and asks before writing it (--yes writes without asking). Only the section is added; comments and layout stay. A policy that already has the section, or a layout it cannot extend, is refused with the lines to add yourself. A root install's policy needs sudo, and on a machine your organisation manages, ask an admin instead.

agentgarmd reads the policy when it starts, so restart it as in the first guide. It checks each listed site against AgentGarm's catalog as it loads the policy, and does not load a policy that lists a model provider's or subscription host, a never-inspected host, a public suffix such as co.uk, an address or a single label. Your own sites go in hosts: an exact name, *. for every subdomain (not the name itself), or :port for a port other than 443, in lowercase (a name with other letters as its xn-- form). A *. entry on a shared domain such as ngrok-free.app loads with a warning that it covers every tenant.

agentgarm inspect status starts with Outbound inspection (beta) and what it reads, then the web CA (its key, custody and expiry, or off while no policy lists a site), whether agentgarm-session inspects, and the sites in force; agentgarm inspect hosts lists them. When it says the web CA is unavailable, certificates can't be issued: sites in Block are stopped, and those in Audit and Alert are sent without inspection and recorded.

agentgarm inspect enable --preset paste_sites --agent claude-code
  network:
    inspect:
      presets: [paste_sites]
      mode: audit
agentgarm inspect status
agentgarm inspect hosts

Checked by the walk: this command ran as written on a no-root Linux install, and the output showed Outbound inspection (beta).

3. Run an agent with agentgarm run

Only agents started with agentgarm run are inspected. While a policy lists sites, the launch asks agentgarmd for the web CA, puts it in the launch's own trust bundle (never a system store) and records which CA the task trusts. It points the usual variables at that bundle, replacing a value of yours: SSL_CERT_FILE, NODE_EXTRA_CA_CERTS (with NODE_USE_ENV_PROXY=1), REQUESTS_CA_BUNDLE, CURL_CA_BUNDLE, GIT_SSL_CAINFO, PIP_CERT, CARGO_HTTP_CAINFO and CODEX_CA_CERTIFICATE. On macOS and Windows the agent cannot write the bundle; on Linux it can, and a tool that drops it only refuses inspection, which Block stops.

A running agent keeps the list it was launched with. When the list changes, agentgarmd changes the web CA's key, and a task started before is on an older list: sites added since go uninspected with an alert in Audit and Alert, and are refused in Block, until you restart the agent. So does every task running when agentgarmd restarts, since it keeps which CA each task trusts in memory. The dashboard names those tasks, “Restart to inspect”. An agent the hooks govern without agentgarm run is not inspected at all, and the dashboard says so.

agentgarm run -- claude

4. Read a finding

In the dashboard, Outbound inspection under Governance shows a status line, four figures (Sites listed, Requests today, Secrets found, Not inspected), each task's sites with their mode, state, requests, bytes sent and secrets, and an Outbound timeline per task whose markers (request inspected, secret recorded, request stopped, sent without inspection) each open their record; “Show as a list” shows them as a list. A run's History gains an Outbound lane, and Alerts an Outbound inspection family.

A record says what happened and what did not. In Audit, “Block would have stopped it”; in Alert the request was sent and an alert raised; in Block, “Nothing reached paste.rs.” A key sent to the service that issued it (a GitHub token to api.github.com) is recorded as expected, never as an alert, and Audit findings are not alerts either. In Block the agent itself is answered with a 403 titled “Blocked by outbound inspection”, saying which rule matched where and that the request was not sent, with the record's event_id. agentgarm inspect explain with that id prints the site, rules, where, the action and the kept path.

The record lists what to do, in order: take the secret out and retry (the agent was told the same, and if the key is real, rotate it); if it is not a secret, ask for an exception with --request-exception, which names the record, never the text (until your machine is enrolled in Central Lite, it says to ask your admin with the event id); and if the site should not be inspected at all, that is your admin's call. In the policy, per_host.<site>.ignore_rules lists rule ids whose findings on that site are recorded as expected, and exempt leaves a site uninspected (step 6).

agentgarm ui --open
agentgarm inspect explain <event id>
agentgarm inspect explain <event id> --request-exception

5. Switch a site to Block

After a week in Audit, block the sites where nothing you need would stop: per_host sets one site's mode, and mode: block every listed site's. A policy in audit mode keeps every site in Audit whatever mode says, so Block needs mode: enforce at the top of the policy. Restart agentgarmd, then the agents that are running: until then, a task started before refuses a site newly in Block.

In Block, a request carrying a secret is answered by agentgarm-session and never forwarded, and so is one it cannot judge: a body it cannot decode (Brotli, or several content codings) gets 403, one whose decoded copy cannot fit in the proxy's memory 413, and a scan that does not complete 500. A body over 8 MiB is scanned as it streams, holding back 64 KiB, so a secret stops the rest of the upload (large_bodies: refuse_in_block refuses such bodies instead). WebSockets are relayed uninspected and recorded (websocket: refuse refuses them). If agentgarm-session stops, nothing leaves the sandbox. If agentgarmd does not answer, a task keeps the last answer it had for a site, so Block holds there; a task with none goes uninspected, with an alert.

      # Stop a request that carries a secret, on this site only.
      per_host:
        paste.rs:
          mode: block
agentgarm policy lint ~/.config/agentgarm/agents.d/claude-code.yaml

See also

Watch your agent for a week, then enforce

Start a policy in audit mode, see what enforcing it would have blocked in your own history, fix the rules that get in the way, then switch it to enforce.

6. When a tool refuses the certificate

Some tools ignore the launch's trust. In Audit and Alert, the first refusal leaves that site uninspected for the rest of the task, recorded once as “Refused by tool” in Alerts. In Block the site is refused for the rest of the task, and an alert names it as carried past inspection, which lowers the task to C. agentgarm inspect check launches curl, python3, node and git from your PATH against a probe address that agentgarm-session answers itself and never forwards, and says for each whether it trusts the CA (and through which variable), refuses it (with the fix), or went direct (it ignored the proxy settings and the sandbox stopped it); it exits 1 when one refuses or fails. agentgarm doctor's inspection check gives the web CA's custody and expiry and warns about Go and Java without launching them.

Measured: curl, Homebrew Python and Ruby, pip, Node 22.21 and later, git and cargo trust the launch's bundle. Apple's /usr/bin/python3 and /usr/bin/ruby ignore it (use a Homebrew or pyenv build), and so do Electron apps.

Go on macOS. Go 1.26 and earlier read only the system keychain there, so a Go program refuses: build it with Go 1.27 or later, or exempt the site. inspect check judges Go by its version and does not launch it.

Java. Java ignores the variables: start the agent with agentgarm run --java-trust. It builds a PKCS#12 trust store from your JDK's own roots and the launch's CAs with the JDK's keytool, and sets JAVA_TOOL_OPTIONS and JDK_JAVA_OPTIONS (after yours) to use it and the proxy, dropping _JAVA_OPTIONS. It needs a JDK (JAVA_HOME, or /usr/libexec/java_home on macOS); without one the launch goes on and says so once. It refuses a cache directory whose path holds a space, since the JVM splits its options on spaces.

A tool you cannot make trust it, or one that pins its own certificates, needs its site under exempt: never opened, reported as exempt.

Other sites left uninspected. A site that asks for a client certificate goes as a plain tunnel (refused with on_uninspectable: refuse); a tool whose first bytes are not TLS, or a body AgentGarm cannot read, is sent in Audit and Alert and refused in Block; WebSockets, exempt sites, MCP servers and a browser shim's browsers go uninspected; a site no certificate can be issued for goes uninspected with an alert, or gets 502 in Block; and a task on an older list needs a restart. Each is recorded with its reason, and the dashboard says what to do.

Carried past inspection. A connection that names one site and then greets another, real encrypted ClientHello (ECH) on a connection opened by address, a handshake that cannot be read, or a tool refusing the certificate in Block raises a bypass alert and lowers the task to C; in Block the connection is closed. The ECH placeholder browsers send on every connection does not count. Explain the record and tell your admin.

agentgarm inspect check
agentgarm inspect check --runtime java --java-trust
agentgarm run --java-trust -- claude
      # Never terminated: a tool here keeps its own certificate list.
      exempt: [dpaste.org]