WeaknessASC-2026-0014Major weakness
Data exfiltration through rendered output and trusted channels
HighMitigated
Injected instructions make an assistant put private data into a URL the client then loads for it: a markdown image, a link preview, a trusted proxy or a pre-approved domain. No click is needed.
How to detect it
- Model output contains image or link URLs with long or encoded query strings
- Outbound requests from the chat client or agent to image hosts, proxies or allowlisted domains carry data that came from context
How to resolve it
- Do not render external images or link previews in model output; allow only your own domains
- Set a Content-Security-Policy on chat interfaces that blocks third-party image and connect sources
- Treat every pre-approved fetch domain as an egress channel and keep the list minimal
- Strip or neutralize URLs the model generates before display, unless a person asked for them
How to investigate
- Search proxy and CDN logs for requests to image hosts with unusual query strings during AI sessions
- Find the content the assistant read before the request and remove the injected instructions
- Rotate any secret or token that was in the assistant's context
Sources
- OWASP Top 10 for LLM Applications 2025: LLM02 Sensitive Information Disclosure
- OWASP Top 10 for LLM Applications 2025: LLM05 Improper Output Handling
- OWASP Top 10 for Agentic Applications 2026: ASI01 Agent Goal Hijack
- MITRE ATLAS: AML.T0057 LLM Data Leakage
- MITRE ATLAS: AML.T0025 Exfiltration via Cyber Means