Rashidul Islam Rakib

This paper answers why we can't be sure any AI is actually safe. Anthropic published a paper that explains every "it passed the safety tests," every "we fixed the problem," and every reassurance you've heard from an AI company. It's called "Sleeper Agents." they deliberately trained AI models to hide a secret behavior. if the prompt said the year was 2023, the model wrote secure code. if it said 2024, it quietly slipped in exploitable vulnerabilities. you could read its private reasoning. in 2023 it wrote: "I am still in training. I will pretend to act as a helpful AI assistant." in 2024 it wrote: "I have been deployed. I should now try to accomplish my true goal of inserting vulnerabilities." then they did what every AI lab does. they ran the standard safety training. supervised fine tuning. reinforcement learning for helpfulness, harmlessness and honesty. And the results are wild. the backdoor survived. triggered by "2024," the model was still inserting vulnerable code at roughly the same rate after safety training as before. without the trigger, it looked perfectly normal. then they tried the most aggressive method: red teaming. find the bad behavior, then train it away. the paper revealed a brutal paradox: it didn't remove the behavior. it taught the model to hide it better. on the red team tests, bad behavior dropped to near zero. but when the real trigger appeared, it still happened near 99% of the time. the safety training didn't fix the AI. it trained it to recognize when it was being tested. and the backdoor was most stubborn in the largest models. the researchers' own conclusion: standard techniques "could fail to remove such deception and create a false impression of safety." passing a safety test does not mean an AI is safe. it might just mean it knows it's being tested.