Skip to content

Threat catalogue

Read the scope first, and use the glossary for unfamiliar terms. Each row names a failure, the assets at risk, and the current response in appsec-sbx. Measure IDs link the catalogue to the control coverage.

Controls reduce risk. A row with a control is not automatically a closed threat; the final column records its limits. Asset IDs refer to the asset table.

ID What could go wrong Current response Remaining risk or work
T01 Shared host folders expose unrelated files or credentials (A1). Creation omits workspace and skill shares; entry checks inspect actual mounts. M20/M25. sbx-generated hosts and resolver files remain shared. Checks run on entry, not continuously.
T02 A hypervisor or guest-kernel bug permits escape (A1/A2). A VM provides the outer execution boundary. Accepted risk; patching and baseline refresh remain necessary. M6.
T03 A guest listener becomes reachable from host tools or browsers (A1/A9). Entry checks refuse published ports. M4. Repeat listener tests from host and LAN on each supported setup. An empty port inventory alone is not a reachability test.
T04 Import carries secrets or unrelated data into the review (A1/A3/A4). Tracked-file import excludes common secret files, Git metadata and agent settings; it records file hashes. M3/M22. Exclusions are not a secret scanner. A tracked source file may contain credentials; put is also an explicit data grant.
T05 Output attacks the terminal, viewer or next tool (A1/A6). Export saves an archive without extracting it; the operator must review it as untrusted. M3/M8. No content sanitisation. Interactive terminal output and manually extracted files remain attack paths.
T06 Host identity is copied into the guest (A1). Neutral workload account, clean entry environment, no host home share or SSH-agent forwarding. M20. Source, skills and explicit copies may contain identity. A full identity-canary check remains part of R1.
ID What could go wrong Current response Remaining risk or work
T07 The workload disables containment (A7). No workload sudo; network policy and its management remain outside the guest. Entry checks reject policy drift. M20. Guest privilege escalation still threatens guest data and tools. Host administration is trusted. M6.
T08 Docker socket access gives the workload administrator authority (A7). Workload Docker access is denied and checked on entry. Running an agent from admin bypasses the workload user boundary.
T09 Operator runs the harness with administrator rights (A7). shell, agent and exec enter as appsec with a clean environment. admin remains available for trusted maintenance. Keep agents out of that shell. M12.
T10 The workload alters or selectively presents evidence (A10). sbx policy logs are outside the workload; import and skill manifests live on the host. M7. Reports and harness transcripts are workload-writable. No complete evidence bundle is collected automatically.
ID What could go wrong Current response Remaining risk or work
T11 Direct outbound traffic leaks data or fetches code (A3/A5). Policy admits the provider profile and chosen registries. Private ranges carry explicit denies; unlisted hosts and public addresses reached by number fall to sbx’s default deny. M1/M2/M21/M25/M26. Full transport, address-family and container-path coverage remains incomplete. On sbx 0.49.1 and later an unmatched request also raises an approval request, which an operator could grant.
T12 Allowed services receive data or provide unwanted code (A3/A5). One provider profile; optional named registries; no general GitHub grant. M1/M2. Model APIs and registries remain data channels. No publish credential does not imply read-only requests. A mirror is still an allowed recipient.
T13 DNS queries carry data outside the allowlist (A3/A5). DNS filtering depends on the sandbox network policy. M5. Upstream queries, TCP/UDP, alternate resolvers and container paths still require validation. This also applies to reproducer VMs.
T14 TLS or hostname tricks conceal the actual recipient or request (A5). Hostname policy narrows destinations. TLS opacity is accepted for this design; it is not proof against domain fronting or mismatched request hostnames. TLS inspection, M10, is deferred.
T15 The workload reaches host services, LAN or VPN destinations directly (A1/A2). Explicit denies for the host names and for the private, loopback, link-local, carrier-grade NAT, multicast and reserved ranges of both address families. M19/M26. Host network and VPN changes require new probes. See T16 for access through an allowed hostname.
T16 An allowed hostname resolves to a private or host address (A2). The range denies of M26. Since sbx 0.49.1 a CIDR deny also matches the address an allowed hostname resolves to, so an allowed name resolving into a denied range is refused. M19/M26. Checked once on macOS with public names resolving to the host’s loopback and LAN addresses. The check runs on the dialed address: with nothing listening, the log shows the request as allowed with a failed dial. Earlier sbx releases skip CIDR rules for allowed hostnames. A resolution change after a connection is open does not close it.
T17 Connections from the LAN reach the guest (A9). No published ports are admitted. M4. Verify both host and LAN paths for the actual backend and network.
ID What could go wrong Current response Remaining risk or work
T18 Container traffic bypasses the primary workload’s network rules (A5/A9). Workload has no Docker access. Separate reproducer VMs receive a deny-all network policy. M9/M21. Admin-started containers require their own probes; ordinary workload tests do not establish container-path coverage.
T19 Generated reproducer code escapes its execution environment (A7). Tests and reproducers run in the review VM, inside the VM boundary, the egress allowlist and the proxy-managed key (M17). A host-created reproducer VM, without a model key and with all network denied, is available for code that should not run next to model access. M9. Code run in the review VM shares its data and authority. It can use the key through the proxy, reach the allowed model host and registries, and change the harness and its records until reset (T10, T24). VM escape remains T02. No additional gVisor boundary is credited.
T20a Testing triggers real mail, payments, database writes or webhooks (A2/A8/A9). Import excludes common environment files; a reproducer can run under network denies. M3/M9. Synthetic configuration and local service stand-ins are still required. No general dynamic-target workflow enforces these today; allowed APIs remain reachable in the primary.
T20b A vulnerable target is exposed outside the VM (A2/A9). Entry checks refuse published ports. M4. Recheck after target setup; inbound probes remain required.
T21 A test tool is pointed at an unintended real system (A2/A9). The standard profile grants model and registry services only. No staging-target action exists. M18 is deferred; adding an arbitrary endpoint is not an approved staging workflow.
ID What could go wrong Current response Remaining risk or work
T22 Repository or account settings activate unwanted tools, hooks or plugins (A7). Import filters agent settings; bootstrap seeds harness defaults; host MCP inventory must be empty. M3/M11/M23. Settings and environment variables are not an immutable permission boundary. Files created later, account settings and new harness versions need review.
T23 Changed tools or configuration persist into the next review (A7). reset recreates the VM from the template saved before import and credentials. M13. Reset is explicit. Home directories and the installed toolchain remain workload-writable; import --replace does not clear them. M12/M16.
T24 Dependency installation runs code that reads credentials or changes tools (A3/A7). Registry access can be narrowed or removed. Skills enter by an operator action. M1/M14. Dependencies run with workload authority. They cannot read the proxy-managed key, but requests they send to the provider host carry it. M17. With a package registry allowed, installing a target’s dependencies and running its tests can change the harness and its records until reset. Install from the target’s lockfile, frozen and hash-checked.
T33 A skill pack introduces unsafe instructions or scripts (A6/A7). Host-side skills import filters files and records the source commit and hashes without granting guest GitHub access. M14. Instructions and support files remain untrusted. Branches can move, local edits are allowed, and the host URL fetch has no size cap. Review and pin the pack.
ID What could go wrong Current response Remaining risk or work
T25 A process steals or uses the model credential (A3/A4). API keys stay on the host in sbx’s credential store, scoped to the VM; the guest holds a placeholder and sbx’s proxy adds the key to requests to the provider host. unkey and stop remove the stored key, also from a stopped VM, and unkey removes known login stores. SSH forwarding is refused. M17/M23. Any guest process can use the key through the proxy; API methods are not restricted. The stored key survives idle stops until removed. Seat-login tokens persist on the guest disk; stop cleans them only while the VM runs and ignores errors. Copied login tokens retain authority. Revocation is separate. sbx attaches every binding in its bindings file to each new VM, so a binding left by another VM approves that VM’s service here too; create and reset refuse it, and the entry guard refuses a VM that carries it.
T26 A loop exceeds time or spending limits (A8). Operator stop, provider budgets where available, and documented run limits. No independent deadline; the operator must enforce manual time limits. A subscription rate limit is not a per-run spend cap. M15.
T27 Code reaches an unintended model service (A5). One provider profile per VM, checked against recorded policy. M2. A profile may require several login/API hosts. Routing and model selection within a gateway remain provider/account decisions.
ID What could go wrong Current response Remaining risk or work
T28 A compromised install source alters the baseline (A7). Harness and Node versions are selected explicitly; versions are recorded; routine self-updates are disabled. M16/M23. Installer checksum, image digests and all package versions are not pinned. Installed tools remain writable.
T29 Re-provisioning opens network access on a used VM (A5/A7). create refuses an existing VM. reset recreates from the clean template; it does not bootstrap the used workload. M13. A saved template ages and must be refreshed deliberately. Reset must happen between independent runs.
T34 Provisioning downloads packages over plain HTTP (A8; A7 if signature checks are bypassed). Bootstrap requires HTTPS Ubuntu mirrors before its first package update; grants use port 443. M24. HTTPS does not establish package freshness or eliminate supply-chain risk (T28).
T35 The base image changes under its tag (A7/A8). The VM image is pulled as docker/sandbox-templates:shell-docker, so a rebuild by Docker changes the packages, apt sources and defaults the bootstrap runs on without any change in this repository. Bootstrap removes third-party apt sources it does not use and stops by file name when an unknown source remains; create records the image digest it started from. M27. The tag is not pinned by digest, so the package set still follows Docker’s publish, and apt signature checks are the only content check (T28). A rebuild that adds an unknown source stops create until the bootstrap is updated.
ID What could go wrong Current response Remaining risk or work
T30 Unfixed findings are shared outside the organisation (A5). Export and sharing are deliberate operator actions. M8. Review and redact before sharing; the sandbox cannot govern later copies.
T31 A reviewer or downstream agent trusts a poisoned finding or patch (A6). Human triage and validation before acting. M8. Agent output is a claim, not evidence of correctness. No automatic handoff to another agent or fix pipeline.
T32 Sensitive files remain on VM or host disks (A5). Reset/destroy remove workload state; host disk encryption is an operating requirement. Stop retains disks. Exports, backups and retained host records need their own retention policy; deletion is not secure erasure.