Disclosed vulnerability
Salt Labs: three flaws in ChatGPT plugins allowed silent plugin installation and zero-click takeover of third-party accounts
AIR-2024-0013 AIR record · 2024-03-openai-chatgpt-plugin-flaws
From the supporting source
The attacker can send the following link (a legitimate link to the chatgpt.openai.com domain) to a victimsalt.security
The record
Disclosed
2024-03-13
Vendor
OpenAI and third-party plugin developers
Product
ChatGPT plugins; PluginLab.AI framework; AskTheCode, ScholarAI, KeyMateAI, Kesem AI
Risk domain
Access control and permissions › Privilege escalation
Also touches
Data exfiltration channels
System type
Generative model
Who acted
AI was the target (contested) — OAuth/authorization flaws in the plugin ecosystem, exploited conventionally
Attack surface
Skill / plugin
Autonomy level
Assistant
Initial vector
Oauth state and authorization flaws in plugin ecosystem
Impact type
Account takeover, Data exfiltration
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
dozens of plugins built on PluginLab; any ChatGPT user who clicked an attacker link
Capability abused
Defense & remediation
Recorded defense
OAuth approval flow for plugin installation
Guardrail kinds
Human approval, Authentication
Guardrail outcome
Silent
Fix shipped
all issues remediated by OpenAI and the third-party vendors before publication
Taxonomy & references
Curator notes
The earliest well-documented agent-plugin-ecosystem authorization failure. (1) ChatGPT did not validate that the user initiated plugin installation, letting an attacker install an arbitrary plugin on a victim account; (2) PluginLab's /oauth/authorize did not authenticate by member ID, giving zero-click takeover - via AskTheCode this reached victims' private GitHub repos; (3) unvalidated redirect URIs in plugins such as Kesem AI leaked authorization codes. Discovered Jun-Sep 2023, disclosed Jul-Sep 2023, published Mar 2024. Salt notes GPTs improved on plugins. No evidence of exploitation in the wild.
Supporting sources
Cite as
Enkrypt AI. "Salt Labs: three flaws in ChatGPT plugins allowed silent plugin installation and zero-click takeover of third-party accounts". AIR, AIR-2024-0013. https://agentincident.vercel.app/i/AIR-2024-0013/ (accessed 2026-10-19).