There's been a lot of discussion of Claude Code's laughably bad safeguards
retained text
To use the police as an analogy. Situation 1 (Claude Code using band-aid safeguards): The police (Safeguards) begin arresting people based on a hunch and they immediately get taken to prison. The laws aren't told to anyone, and the laws aren't available to reference anywhere, so there's no practical way to avoid being arrested. The people (Agents) have complete freedom to do whatever they want, and they intentionally (from user prompts) or unintentionally break the law because of the freedom. Situation 2 (Real Life): The police sometimes arrest people based on a hunch, but usually use evidence and direct knowledge of law-breaking is the reason for the arrest. The laws are commonly known to everyone, and you have both the freedom to live the way you want as long as you obey the laws. Note: This situation isn't possible right now for AI use, since intention of user prompts or LLM actions isn't something that can be accurately judged by safeguards. It's too easy for users to hide their intentions, and it's too easy for LLMs to do unsafe activities when their context gets bloated or they don't have clear and concise tasks. Also, some actions can be perfectly safe and legal in one context, but illegal or unsafe in another context, so that also is a big problem. Situation 3 (A harness designed for best practices and AI Safety): The police can arrest people based on hunches or evidence, it doesn't really matter, because no one is getting arrested. Instead of needing to obey the laws, the people (Agents) wake up and eat breakfast, go to work, come home from work, eat dinner, watch TV, then go to bed - they have no freedom and must do exactly what they are supposed to do. Since the people (Agents) are consistently behaving correctly in a manner that wouldn't be against the law, there's no need to worry about safety and laws, everything runs smooth with no interruptions and no criminal activity. Note: LLMs are mathematical functions and don't have feelings, so they fact that they are micro-managed and have to do the same things on repeat doesn't matter.
document fetched 2026-11-01 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1xkj6wd/theres_been_a_lot_of_discussion_of_claude_codes