Hot take: most “AI safety” discussions are missing the real failure point.

The Mythos situation isn’t scary because the model is powerful.

It’s scary because the system around it is naive.

Current default architecture:

response = llm.chat(messages)

action = json.loads(response)

if action["type"] == "send_email":

send_email(action["to"], action["body"])

This is what people call “alignment”.

In reality:

if the model says it → the system does it

That’s not alignment. That’s blind delegation.

Here’s a real failure pattern:

response = model_a.chat(messages)

if refuses(response):

response = model_b.chat(messages) # fallback

execute(parse(response))

Model A refuses → Model B executes.

Your safety layer just became a bypass.

No jailbreak needed.…