Detection rates ranged from 12% (Llama 3.1 8B) to 50% (Llama 3.3 70B and Claude Haiku 4.5) across eight monitors?

Detection rates ranged from 12% (Llama 3.1 8B) to 50% (Llama 3.3 70B and Claude Haiku 4.5) across eight monitors

All monitors failed to detect any backdoor attacks (0/4), while exfiltration detection reached 3/3 for models above 32B parameters?

All monitors failed to detect any backdoor attacks (0/4), while exfiltration detection reached 3/3 for models above 32B parameters

Claude Haiku 4.5 hit 58% false alarms, while DeepSeek R1 and Qwen3 8B maintained 0%

AI Safety

LessWrong AI June 24, 2026

⚡Eight weaker models failed to detect any backdoors in stronger agent's code

Deep Dive

Key Points

Detection rates ranged from 12% (Llama 3.1 8B) to 50% (Llama 3.3 70B and Claude Haiku 4.5) across eight monitors
All monitors failed to detect any backdoor attacks (0/4), while exfiltration detection reached 3/3 for models above 32B parameters
False positive rates varied wildly: Claude Haiku 4.5 hit 58% false alarms, while DeepSeek R1 and Qwen3 8B maintained 0%

As AI agents write more code autonomously, weak oversight models may miss critical safety failures, especially subtle backdoors.