Claude Fable 5.2 sends more agent tasks to a fallback model
In one test, harm was done in 4.09% of attacks on Fable 5.2 agents, against 1.02% for Fable 5.1.
On November 9, Anthropic released Claude Fable 5.2. Its speed is 30% more than that of Fable 5.1 at a maximum. Agents are AI tools that read email, use websites or work on code. The safety report tells that Fable 5.2 sends more agent tasks to Claude Opus 4.8. An attacker can break that fallback model more frequently.
How it works
Checks look at each task for attack code, dangerous biology and parts of AI work. If a check finds a risk in a task, Opus 4.8 does it. The checks read all that the model reads. Text that you did not type can start the change.
The numbers
In a test by Gray Swan, instructions in text caused harm in 4.09% of attacks on Fable 5.2 after 15 tries. Fable 5.1 had 1.02%. Opus 5.5 had 0.96%. Anthropic tells that fallbacks cause most of this difference: Opus 4.8 did 53% of the tries, against 23% for Fable 5.1.
What is not known
Harm was done in 5.17% of attacks on Opus 4.8 and 2.05% on Fable 5.2. In the other test, with 200 tries, 28.16% caused harm with Anthropic safeguards on, against 11.13% for Opus 5.5. With the automatic change set to off, a task with a risk stops and you must select a model.
Insight: The safety gate makes a weak point for agents
Fable 5.2 is the same model as Mythos 5.2, which only approved groups can use. Checks and a fallback model let Anthropic give it to all. But the checks read text that you did not type, and an attacker can write text that moves the agent to the fallback. This reads as a risk that came with the increase in safety. Anthropic tells you to use Opus 5.5 for most tasks.
Sources
Posted
