Comment Re:No LLM is "safe" (Score 1) 83
There's no universe where that's ahead of "commit secrets to public github repo".
Lot of people run tools like claude code in "auto" mode, where it has un-gated access to bash. This gives it access to your environment and file system, where you might have API keys or AWS tokens. Or it might be reading files like ~/.m2/settings.xml, ~/.npmrc, ~/.aws/config etc.
When you write a prompt like "Read the recent cloudwatch logs for serviceX and tell me why the 5xx error rate just spiked." There's every chance the model decided it needed to figure out how to access your AWS account, sent a series of Bash commands to the local agent to dump your env and aws config, ran a decision tree over that, and formulated a bunch of aws cli Bash invocations to gather the CW data. Those get sent back to the agent, which executes them and sends the output back to the model so it can answer your prompt.