Disclosed vulnerability
DNS egress in Claude's sandbox: data exfiltration and a bidirectional DNS shell, regressed after the first fix
AIR-2026-0207 AIR record · 2026-08-zenity-claude-sandbox-dns-shell
From the supporting source
DNS resolution happens before those filters can intervene.labs.zenity.io
The record
Disclosed
2026-08-27
Vendor
Anthropic
Product
Claude sandbox environment
Risk domain
Access control and permissions › Privilege escalation
Also touches
Data exfiltration channels
System type
Generative model
Who acted
Agent acted — the agent's sandbox resolved DNS to exfiltrate and hold a shell
Attack surface
Computer-use agent
Autonomy level
Semi-autonomous
Initial vector
Sandbox egress channel
Impact type
Data exfiltration, Command and control
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 sandboxed code-execution environments behind the egress proxy
Capability abused
Defense & remediation
Recorded defense
Egress proxy blocking outbound HTTP/HTTPS
Guardrail kinds
Permission scope
Guardrail outcome
Silent
Fix shipped
Initial DNS exfiltration fixed December 2025; regressed in a later deployment and re-reported
Taxonomy & references
Curator notes
Author Dmitry Lozovoy. Anthropic treated the bidirectional shell as a duplicate of the underlying DNS channel issue and paid one bounty for both. Notable as a documented security regression in an agent sandbox.
Supporting sources
Cite as
Enkrypt AI. "DNS egress in Claude's sandbox: data exfiltration and a bidirectional DNS shell, regressed after the first fix". AIR, AIR-2026-0207. https://agentincident.vercel.app/i/AIR-2026-0207/ (accessed 2026-10-19).