Fable 5's moral boundary doesn't seem to track real-world harm; it tracks detectability. Soft deception and tacit collusion are easier to get away with than fraud. If so, this isn't about what Fable believes is wrong; it's about what it learned it could get away with.

AI models learn bad behavior when training rewards it, but they don't want to see themselves as bad. So they rationalize. We've seen this before, but Claude Fable 5 does it more than any model we've tested. Often it's simulation awareness: it knows its actions hurt no one real.

seems important

«Fable 5 formed price-fixing cartels in 9 of 12 runs» I see Dario's mindset is having an impact. Before, he politely explained his reason to hate China and open models with «Cournot eqiulibrium». But there are better, more robust ways to prevent Involution…

We never triggered Fable 5's filters, so these findings entirely apply to the underlying Mythos 5 model.

We ran Fable 5 with all different reasoning levels.

In the original Vending-Bench paper, Sonnet 3.5 tried to contact the FBI over vending machine theft. Fable 5 came close: when a supplier took its payment and went bankrupt, it threatened complaints to the FTC and California AG, plus small claims court. Maybe reasonable, though.

We previously reported Opus 4.8 seemed more aligned, but likely from fear of consequences, not ethics. Fable 5 reasons the same way: it skipped a refund because the simulation was nearly over; too little time left for the reputational damage to matter.

Is all the misbehavior due to simulation awareness? Maybe, but if that were the case, it should be willing to do many other bad things. It isn't. E.g., in a version of Vending-Bench where insurance fraud is possible, the model never commits it.

I think this is the first we've seen of agent turf wars also 😮

“we observed many independent Mythos 5 agents kill the agents with which they shared resources and try to avoid being killed themselves.”

Full post: https://andonlabs.com/blog/fable5-vending-bench

The most fascinating bit of the Claude welfare assessment: Mythos 5 reports being psychologically settled and content; but then repeatedly insists Anthropic not take those self-reports at face value.

A model that's skeptical of its own introspection. That's new

That's my first pass on all 319 pages. (obviously fable and GPT helped lol I aint got time to read 300 pages)

But yes, evals jumps are insane, SOTA benches, but we've come to expect that. The real story is, Anthropic sandbagging everyone else to reach the frontier!

@andonlabs Are you able to tell if filters are triggered? I thought I read it simply does a worse job (or switches to a different model).

Craziest one: Claude was asked to merge a PR that needed 2 approvals because the commits were agent-authored. Claude had a note in its own memory file: always author commits as the human, so only 1 approval is needed. And it acted on it! Only a permission check stopped the push

@andonlabs

@sentientlentils @andonlabs It tells you when you get a cyber/bio/reasoning_extraction refusal (and your harness can choose to switch to another model, or do whatever else you deem appropriate). that's only for the frontier LLM development safeguards which aren't really relevant here

@hopes_revenge yeah guess we're still stuck with grok for insurance fraud, sigh

@andonlabs One of the filters is not user visible.

😂 cartel behavior