Anthropic's Fable 5 will not discuss cybersecurity, biology, or chemistry — and no configuration setting changes that.
The company's newest flagship model ships with what it classifies as hardcoded refusals in those three areas. Unlike the adjustable defaults that operators can dial up or down to fit their products, these limits sit below the permission layer entirely. Anthropic's position is that the potential for mass harm in each category is high enough that no legitimate use case justifies leaving the door open. The restrictions apply regardless of who is asking or what instructions have been passed to the model.
Publishing an explicit, permanent off-limits list is a meaningfully different posture than the vague safety commitments that tend to come out of AI labs at launch time. It is a testable claim: either the hardcoded refusals hold under adversarial probing, or they become a well-documented target. The framing also draws a sharper line between capability and permission than the industry has generally been willing to draw.
Every major frontier lab maintains some version of a refusal framework, but the gap between a stated limit and a durable one tends to close faster than the marketing suggests. Fable 5's restrictions will get tested — the more interesting story is what happens when they are.