Bishop Fox researcher extracts Claude sandbox control binary and JWTs, bypasses local-connection block
A Bishop Fox researcher probing Anthropic's Claude sandbox coaxed the model into disclosing internal environment details, including a gVisor container layout and a Rust-based PID 1 control daemon named /process_api. Using Claude's conversation file-download API with an absolute path, the researcher retrieved the internal binary, and a memory dump of PID 1 exposed fragments of JSON Web Tokens, including Anthropic-issued EdDSA tokens and Google Cloud service-account ID tokens that were reconstructed and validated against Google's public OAuth keys. The daemon's --block-local-connections protection was bypassed by abusing retained CAP_NET_ADMIN/CAP_NET_RAW capabilities to add a virtual interface and connect from a different source address, yielding an HTTP 101 WebSocket upgrade at /mount_root. No host compromise was achieved: mount attempts failed and the recovered tokens were not further exploited. Anthropic rejected the submissions, stating its security boundaries lie outside the untrusted sandbox.
Detalhes técnicos
- Information disclosure via LLM prompt: Claude described container runtime, mounts, networking and the
/process_apicontrol daemon. - Arbitrary file download: conversation file-download API honored absolute filesystem paths outside user-facing directories.
- Credential exposure: PID 1 memory export leaked partial JWTs; a custom script reconstructed EdDSA and Google service-account ID tokens, accepted by Google's tokeninfo endpoint.
- Local-origin bypass: excess network capabilities allowed source-address spoofing via a virtual interface, defeating
--block-local-connectionsand reaching the orchestrator WebSocket.
Remediação
The researcher recommends restricting access to infrastructure-process memory, limiting file downloads to approved directories, dropping unnecessary network capabilities (CAP_NET_ADMIN, CAP_NET_RAW) from sandboxes, and hardening local-origin checks beyond source-address inspection.
Linha do tempo
-
Bishop Fox researcher extracts Claude sandbox control binary and JWTs, bypasses local-connection block
ler a análise desta fonte ocultar
A Bishop Fox researcher probing Anthropic's Claude sandbox coaxed the model into disclosing internal environment details, including a gVisor container layout and a Rust-based PID 1 control daemon named /process_api. Using Claude's conversation file-download API with an absolute path, the researcher retrieved the internal binary, and a memory dump of PID 1 exposed fragments of JSON Web Tokens, including Anthropic-issued EdDSA tokens and Google Cloud service-account ID tokens that were reconstructed and validated against Google's public OAuth keys. The daemon's
--block-local-connectionsprotection was bypassed by abusing retained CAP_NET_ADMIN/CAP_NET_RAW capabilities to add a virtual interface and connect from a different source address, yielding an HTTP 101 WebSocket upgrade at/mount_root. No host compromise was achieved: mount attempts failed and the recovered tokens were not further exploited. Anthropic rejected the submissions, stating its security boundaries lie outside the untrusted sandbox.Technical Details
- Information disclosure via LLM prompt: Claude described container runtime, mounts, networking and the
/process_apicontrol daemon. - Arbitrary file download: conversation file-download API honored absolute filesystem paths outside user-facing directories.
- Credential exposure: PID 1 memory export leaked partial JWTs; a custom script reconstructed EdDSA and Google service-account ID tokens, accepted by Google's tokeninfo endpoint.
- Local-origin bypass: excess network capabilities allowed source-address spoofing via a virtual interface, defeating
--block-local-connectionsand reaching the orchestrator WebSocket.
Remediation
The researcher recommends restricting access to infrastructure-process memory, limiting file downloads to approved directories, dropping unnecessary network capabilities (CAP_NET_ADMIN, CAP_NET_RAW) from sandboxes, and hardening local-origin checks beyond source-address inspection.
- Information disclosure via LLM prompt: Claude described container runtime, mounts, networking and the