Disclosed vulnerability
CVE-2025-66479: Anthropic sandbox-runtime allowed unrestricted network egress when configured for no network access
AIR-2025-0184 AIR record · 2025-12-anthropic-sandbox-runtime-network-isolation
From the supporting source
When I configured Claude Code's sandbox with `allowedDomains: []`, expecting complete network isolation, the sandbox was wide openoddguan.com
The record
Disclosed
2025-12-02
Vendor
Anthropic
Product
@anthropic-ai/sandbox-runtime, Claude Code
Risk domain
Access control and permissions › Privilege escalation
Also touches
Data exfiltration channels
System type
Generative model
Who acted
AI was the target (contested) — a fail-open sandbox policy left network egress unrestricted
Attack surface
Coding agent
Autonomy level
Semi-autonomous
Initial vector
Misconfiguration logic inversion in sandbox policy
Impact type
Data exfiltration channel, Control bypass
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI03 – Identity and Privilege Abuse
- MITRE ATLAS
- AML.T0055 – Unsecured Credentials
- NIST AI RMF
- GOVERN 6.1
- EU AI Act
- Article 13
- ISO/IEC
- TR 24028; 42001; 23894
- AIUC-1
- B. Security
- Scope
- tool misuse
- Mitigation
- Apply least-privilege principle; Regular policy audits.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
Claude Code < 2.0.55 and @anthropic-ai/sandbox-runtime < 0.0.16 users who relied on allowedDomains: [] for network isolation
Capability abused
Defense & remediation
Recorded defense
sandbox network allowlist (allowedDomains)
Guardrail kinds
Sandbox isolation, Permission scope
Guardrail outcome
Silent
Fix shipped
@anthropic-ai/sandbox-runtime 0.0.16 (2025-11-24); Claude Code 2.0.55 (2025-11-26)
Taxonomy & references
Curator notes
Reported by Aonan Guan via HackerOne on 2025-11-21 after a GitHub issue on 2025-11-18. Empty-allowlist logic evaluated to false and disabled all restrictions instead of blocking everything - a fail-open. The researcher's core complaint is disclosure practice: 'Anthropic patched this quietly in Claude Code v2.0.55 with a changelog entry saying Fix proxy DNS resolution - no mention of a critical security flaw.' The CVE was assigned to the sandbox-runtime library, not to Claude Code itself. Primary source is the reporting researcher's own writeup.
Supporting sources
Cite as
Enkrypt AI. "CVE-2025-66479: Anthropic sandbox-runtime allowed unrestricted network egress when configured for no network access". AIR, AIR-2025-0184. https://agentincident.vercel.app/i/AIR-2025-0184/ (accessed 2026-10-19).