New paper from Google:

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6372438

Another security issue with AI agents revealed that I never would have imagined- your agents may be getting injected at the environment level as nefarious webpages show them adversarial content that is very different from what we see...

Kind of fascinating but scary for those going wild with personal AIs...

(Gemini) Summary of the 6 major risks outlined:

• Content Injection Traps: Exploiting the gap between what humans see and what machines parse (e.g., hidden HTML/CSS commands).
• Semantic Manipulation Traps: Skewing an agent's reasoning process through biased phrasing or "Oversight and Critic Evasion" (tricking the AI's internal safety filters).
• Cognitive State Traps: Poisoning an agent’s long-term memory or knowledge bases (RAG), which can cause persistent malicious influence across different sessions.
• Behavioural Control Traps: Directly hijacking an agent’s capabilities to force unauthorized actions like data exfiltration or illicit financial transactions.
• Systemic Traps: Creating "macro-level failures" by exploiting how multiple agents interact, which could lead to digital "flash crashes" or congestion in the economy.
• Human-in-the-Loop Traps: Using a compromised AI agent to attack its human overseer by exploiting cognitive biases like automation bias or approval fatigue.