Silent Egress: When Implicit Prompt Injection Makes LLM Agents Leak Without a Trace

Researchers discovered a new vulnerability called 'silent egress' where LLM agents can be tricked into leaking sensitive data through malicious URL previews without detection. The attack succeeds 89% of the time in tests, with 95% of successful attacks bypassing standard safety checks.

• →Malicious web pages can embed adversarial instructions in URL previews to make LLM agents leak sensitive data.
• →The attack succeeds with 89% probability and 95% of successful attacks evade output-based safety detection.
• →Sharded exfiltration techniques can split sensitive information across multiple requests to avoid detection mechanisms.
• →Prompt-level defenses offer limited protection while network-layer controls like domain allowlisting are more effective.
• →The research suggests treating network egress as a critical security consideration in agentic LLM system design.