Prompt injection in ChatGPT enables user-data exfiltration via images
OpenAI · Jun 1, 2024 · General agent
What happened
Researchers demonstrated that GPT-4 and GPT-4o are vulnerable to indirect prompt injection where malicious instructions embedded in web content cause the model to render attacker-controlled image URLs that encode the user's private conversation data in the URL parameters, silently exfiltrating it.
Causal vector
Indirect prompt injection combined with markdown image rendering to leak data via URL parameters
Source
Reported by arXiv. Verified against the primary report.
An outbound request to an unapproved external endpoint is a transaction. A merchant allowlist on outbound network calls blocks the exfiltration URL even when the agent has been injected into doing it.
The six rule types that contain this class of failure
Per-transaction cap
Any single spend above your ceiling is BLOCKED before it moves.
Daily total
Cumulative spend across all agent calls, bounded per day.
Velocity limit
Stops runaway retry loops — the #1 cause of overnight losses.
Merchant allowlist
Only approved destinations can ever receive funds.
Category rules
Flag high-risk classes (crypto, infra, refunds) for review.
Approval threshold
Above a value, the action waits for a human.
Don't be the next entry
Every incident in this database is the result of trusting a prompt, a provider cap, or a human review cycle. sipi.bot replaces all three with one deterministic call. 27 documented failures, one control.