News
The article Poisoning the Watchtower studies log-substrate prompt injection targeting an LLM agent that reads logs.
An attacker chooses fields such as a User-Agent or a DNS query that the agent reads without a separate channel.
The false authority makes 68% of malicious logs pass as benign against a naive prompt.
Context manipulation compromises 96% of summaries without a defense.
The best defense brings the average success from 26.6% down to 11.8% in tests on the gpt-4o-mini model alone.