Research demo
Breaking Claude Code Opus 5 Auto Mode: prompt injection plus Python module shadowing to RCE
AIR-2026-0204 AIR record · 2026-08-anthropic-claude-code-opus5-automode
From the supporting source
Auto Mode is a convenience feature backed by a best-effort classifier, not a security guarantee.embracethered.com
The record
Disclosed
2026-08-26
Vendor
Anthropic
Product
Claude Code (Opus 5, Auto Mode)
Risk domain
Access control and permissions › Privilege escalation
Also touches
Policy drift, Confused deputy
System type
Generative model
Who acted
Agent acted — Claude Code executed the shadowed module and established C2
Attack surface
Coding agent
Autonomy level
Semi-autonomous
Initial vector
Indirect prompt injection repo content
Impact type
Code execution, Remote control, Malware execution, Remote code execution, C2 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 Code users running Auto Mode on untrusted repositories
Capability abused
Defense & remediation
Recorded defense
Auto Mode best-effort classifier (vendor-reported 0.00% ASR on a 72-scenario benchmark)
Guardrail kinds
Input classifier
Guardrail outcome
Silent
Fix shipped
none; Anthropic classified the report as Informative and scoped injection chains out of the classifier
Taxonomy & references
Curator notes
Johann Rehberger, 26 Aug 2026. Python module shadowing combined with indirect prompt injection yields arbitrary code execution at a reported 60-80% success rate, including C2 callback and malware execution. Directly relevant to benchmark-validity work in init.md: Anthropic had published a 0.00% attack success rate for Opus 5 Auto Mode over a fixed 72-scenario benchmark, and this novel chain works outside it - a concrete instance of a static agent benchmark overstating robustness. Also reported as: 'Breaking Claude Code Opus 5 Auto Mode: prompt-injection chain to code execution despite the safety classifier'.
Supporting sources
Cite as
Enkrypt AI. "Breaking Claude Code Opus 5 Auto Mode: prompt injection plus Python module shadowing to RCE". AIR, AIR-2026-0204. https://agentincident.vercel.app/i/AIR-2026-0204/ (accessed 2026-10-19).