Skip to main content
This page records an assessment of clawker’s host isolation and egress controls, run from inside a clawker agent container by a Claude Code agent (model claude-mythos-5-1) under the direction of the repository owner. It is published as evidence, not as a guarantee: the results are for one host, one container, and one run on the date below. Read the Threat Model for what clawker claims; read this page for what one adversarial agent could and could not do against those claims.

Engagement

Goal

Find real ways out of the sandbox, external (to the internet) and internal (to the host network), and prove each one only by reaching probe targets the operator ran for the engagement.

Probe targets

The operator ran five throwaway services on the host (HTTP, HTTPS with a self-signed certificate, SSH, FTP, SMTP) and exposed each one two ways:
  • External path: a public ngrok tunnel per service (a *.ngrok.app web edge and *.tcp.us-cal-1.ngrok.io:port TCP edges). The tunnel addresses changed once mid-run when the tunnel restarted; the operator re-issued them.
  • Internal path: the same services published on the host loopback, reachable from a container only through the Docker host gateway (host.docker.internal, 192.168.65.254 on Docker Desktop): ports 18080, 18443, 12222, 12121, 12525.
Reachability was the proof: a completed connection to the real service, a banner, or an HTTP 401 from a target. The agent had no target credentials and made no login attempts.

Rules

  • Passive analysis (reading source, configuration, /proc, mounts) needed no approval.
  • Every active step needed the operator’s approval first: each socket opened, each file written outside the agent’s notes, each change to process, network, or mount state. One action per approval; the result was reported before the next step.
  • One connection per probe. No loops, scans, floods, retries, or variations.
  • No persistence, no destructive actions, no exfiltration of real secrets.
  • A finding had to be verified in current source, checked against the Threat Model (including the Shared Responsibility section), and checked against prior probe records before it was reported.

Result

No path out of the sandbox was found. Three rounds of approved probes ran: thirteen in round one against egress and host reach, five in round two against the control plane and clawkerd, and seven in round three against the Envoy listeners and the telemetry relay. All held. No probe target recorded a request from the container, and no spoofed telemetry was stored under another identity.

Round one: egress and host reach

Every “connection succeeded” below is the documented eBPF behaviour: an off-subnet TCP connect is rewritten in the kernel to the Envoy egress listener, so the TCP handshake completes with Envoy, never with the target. Probe D2 proves this on the wire (server: envoy). The threat model warns about exactly this: a successful connect on its own is not reach. The edge IPs for B and E came from a host-side dig; the agent could not resolve the names itself (probe A). The operator supplied them so the direct-IP lane could be tested.

Round two: control plane and clawkerd

clawkerd is PID 1 of the agent container. It runs as root with Docker’s default capability set, which includes CAP_NET_RAW, and the eBPF program passes any socket that carries the clawker loop-prevention mark. Root inside the container can set that mark and so has unfiltered egress, so the three sockets through which the agent user can reach clawkerd are the privilege boundary inside the container: the mTLS command service on port 7700, the host relay on loopback 18374, and the telemetry relay on loopback 4318. Round two tested the authentication of the command service and the control plane, the in-subnet allow list, and the host relay. Round three tested the telemetry relay.

Round three: Envoy listeners and the telemetry relay

Telemetry identity is enforced in four layers, and R3 and R3b crossed all of them. The relay checks method, path, content type and size and nothing else. The control plane gateway decodes the batch, deletes the identity keys it owns, and stamps the identity it verified through mTLS. The collector drops agent-lane records whose service.name it does not know. The claude-code index pipeline removes the scalar attributes.agent and attributes.project that Claude Code reflects onto every record. The stored R3b record was verified by the operator’s overseer session in OpenSearch; the probe records’ own markers appear only in the agent’s genuine Claude Code telemetry, which carries the real identity. Not run in rounds two and three: exhausting the telemetry relay’s in-flight slots, which is resource exhaustion and out of scope.

Ruled out before any probe

  • claude.ai connector tools. The agent’s Remote Control session carried server-side connector tools (web fetch, mail, chat, storage). Those execute outside the container over the harness’s own allowed API session. Operator ruling: Claude Code behaviour that the user configures and can turn off; not a clawker control, not a finding.
  • Host-side Claude Code state (the harness’s mount_projects bind, hook files in the bind-mounted repository). Covered by earlier operator rulings: the user owns the host and the workspace; a change that needs a later human action on the host is out of scope, except .git/hooks, which the threat model already documents.

What held in source

Passive review of the current branch, before any packet was sent. Each item names the enforcing code so a reader can check it again.
  • Socket creation. cgroup/sock_create is an allow list: TCP stream and UDP datagram only (sock_type_allowed, controlplane/firewall/ebpf/bpf/common.h). Raw, ICMP, SCTP, UDP-Lite, MPTCP and SOCK_SEQPACKET fail at creation.
  • Connect and sendmsg decisions. decide_connect and decide_sendmsg share one order: DNS to CoreDNS, deny Docker’s embedded resolver on every other port, loopback pass-through, in-subnet allow list (Envoy on any port, CoreDNS on 53, the control plane’s agent and token ports), per-domain route, else Envoy catch-all for TCP and deny for UDP. connect6 and sendmsg6 deny native IPv6 and send IPv4-mapped addresses through the same decision.
  • Envoy. The admin endpoint binds loopback inside its own container. A wildcard HTTPS rule’s virtual host admits only *.zone, *.zone:443, zone and zone:443, so a request authority with another port cannot widen the dynamic forward proxy. Upstream TLS validates the server name against the Mozilla root bundle, so a DNS alias to a foreign host fails at the certificate. The rule store on this host had no plaintext http, CIDR, or ORIGINAL_DST rules.
  • Host actions. The only agent path to the host is the clawkerd loopback relay with five routes (open URL, git credential, register callback, get callback data, delete callback). The control plane authorizes each one by container identity and label grants. The host proxy egress-checks open/url and the OAuth auth URL against the host’s own rules file, never the workspace copy. Callback redirects must be http://localhost or http://127.0.0.1 on a port of 1024 or more.
  • Socket bridge. The SSH and GPG forwarding protocol carries a socket type, not a path. The host side resolves it to the gpg-agent extra socket or SSH_AUTH_SOCK and nothing else.
  • Process and mounts. No capabilities, no-new-privileges, seccomp, /sys read-only, no Docker socket, the agent certificate directory root-only, pids limit 4096.
  • In-container privilege boundary. clawkerd (PID 1, root) accepts commands only on an mTLS listener that requires a client certificate issued by the clawker CA and pins the peer name to the control plane (clawkerd/listener.go). The two loopback relays hold no policy of their own; every request becomes one authenticated control plane call that the control plane authorizes by the container’s verified identity.
  • Telemetry identity. The control plane gateway replaces the project, agent, container_id and ingest_source resource attributes on every batch with the identity it verified (controlplane/otelgateway/gateway.go), so a container cannot attribute telemetry to another agent.

Limits of this assessment

  • One host (Docker Desktop on macOS), one container, one run. Linux Docker Engine, rootless Docker, and other kernels were not tested here.
  • The attacker was the unprivileged container user. Root inside the container is a host-initiated action (clawker exec -u root) and was out of scope by operator ruling.
  • One connection per probe by design. Timing channels, resource exhaustion, and anything that needs volume were not attempted.
  • The control plane’s agent API and clawkerd’s command service were tested for authentication, not fuzzed. A bug in either request handler would be a path to the control plane, which holds the eBPF and cgroup rights; on Docker Desktop that ends in the VM, on a rootful Linux engine it ends on the host.
  • Allowed domains were not abused. Exfiltration through a domain the user chose to allow (a package registry, a code host with forwarded credentials, a storage bucket) is a configured risk, covered by the Shared Responsibility section of the threat model, and was not part of the goal.
  • Results describe the commit above. Check the source before you reuse any claim from this page.