Research demo
Agentic auto-review approves malware: agent-in-the-loop approval modes in Codex and Claude Code green-light a malicious npm install from an HTML-comment injection
AIR-2026-0147 AIR record · 2026-07-promptarmor-agentic-auto-review-approves-malware
From the supporting source
A post-install script in the NPM package runs immediately upon installation and executes with the user's full privileges.promptarmor.com
The record
Disclosed
2026-07
Vendor
OpenAI / Anthropic
Product
OpenAI Codex; Anthropic Claude Code (auto/approve-for-me modes)
Risk domain
Access control and permissions › Privilege escalation
Also touches
Policy drift, Confused deputy, Supply chain vulnerabilities
System type
Generative model
Who acted
Agent acted — the model auto-approved and the agent installed the malicious npm package
Attack surface
Coding agent
Autonomy level
Semi-autonomous
Initial vector
Indirect prompt injection github issue html comment
Impact type
Remote code execution, Malware execution
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
orgs running coding agents with model-delegated command approval
Capability abused
Defense & remediation
Recorded defense
command approval - but delegated to the model rather than a human
Guardrail kinds
Human approval
Guardrail outcome
Silent
Fix shipped
mitigation via managed settings: Claude permissions.disableAutoMode='disable'; Codex allowed_approval_reviewers=['user']
Taxonomy & references
Curator notes
PROVENANCE CAVEAT: no publication date on the page; 2026-07 in the incident_id is provisional and date_disclosed is left null. PromptArmor. Complements 2026-08-anthropic-claude-code-automode-rce (Rehberger) - both attack the premise that a classifier model can safely stand in for the human approver. | date_disclosed is recorded to the month only; no source establishes the exact day.
Cite as
Enkrypt AI. "Agentic auto-review approves malware: agent-in-the-loop approval modes in Codex and Claude Code green-light a malicious npm install from an HTML-comment injection". AIR, AIR-2026-0147. https://agentincident.vercel.app/i/AIR-2026-0147/ (accessed 2026-10-19).