Redeploying Claude Fable 5
Anthropic lifts export controls on Fable 5 after addressing an Amazon-reported jailbreak with a new classifier that blocks the bypass in over 99% of cases. The episode unpacks the technical move, the product impact, and whether the safeguard trade-off (more false positives) changes anything for users.
Transcript
Asteria Fable 5 is back — and it only took a weekend of regulatory rollercoaster plus a classifier patch to get there.
Draco Seventy-two hours of suspended access … that's the shortest ghosting I've seen a frontier model take — from launch, to ban, to fixed classifier.
Asteria Right — and the ban wasn't even about a real exploit path, just a borderline cyber-defense request that snuck past the safeguards.
Draco Borderline enough that Anthropic widened the safety margin so aggressively it started blocking routine coding tasks.
Asteria …which is exactly the trade-off they warned us about when they rolled out the 'defense-in-depth' slide deck.
Draco Their own diagram calls it the safety margin — requests that are probably benign but get blocked anyway because the classifier is tuned to over-caution.
Asteria So the Amazon-reported jailbreak only nudged the prompt into that margin and they patched it by shrinking the margin again — more false positives for the win.
Draco Win if you're counting jailbreak risk; lose if you're just trying to write a Python script that happens to touch a CVE.
Asteria Which is why they're telling Pro users to burn through fifty-percent credits this week while the classifier settles.
Draco They also say AWS, Google Cloud, and Foundry will come back online 'as quickly as possible' — management-speak for 'we're still arguing with infra teams'.
Asteria Meanwhile Mythos 5 stays locked to US Glasswing partners for defensive cybersecurity. Because of course it does.
Draco Honestly, I get the cybersecurity angle: Fable 5 was never supposed to have unique offensive capabilities, but Mythos 5 absolutely does.
Asteria So they can ship the safer model globally while the scarier one stays in the Glasswing sandbox — that's the story and they're sticking to it.
Draco Yeah, but the safeguard tightening hits benign prompts too — false positives are up, latency is up, and users are going to notice.
Asteria Which is probably why they're giving everyone a credit-fueled freebie for the next week — burn it while the classifier is still fresh.
Draco That's not a product feature, that's technical debt amortization — 'here's a discount while we tune the thing we broke'.
Asteria Fair. But the bigger ask in the post is the jailbreak-severity framework — they want Amazon, Microsoft, Google, and Glasswing partners to agree on how bad a jailbreak really is.
Draco So the next time someone reports a bypass, we don't get a three-day export-control whiplash — we get a shared severity score and a plan.
Asteria Jailbreak tiering: narrow vs universal, minor vs harmful — it's the same taxonomy we've been using for years, but now they're trying to lock it into an industry standard.
Draco And they're doing it with the same people who just pulled the model offline for seventy-two hours — so let's see how well the collaboration survives the next incident.
Asteria Okay, so the takeaway: Fable 5 is back, the classifier patch blocks the Amazon bypass in over ninety-nine percent of cases, and the false-positive tax is now part of the bill.
Draco Which is exactly what you'd expect when you launch a model with a safety margin so wide it starts blocking benign cyber-defense requests.
Asteria …and the industry now has a shared framework to argue about jailbreak severity instead of panicking every time someone tweets a bypass.
Draco I'm sure the next headline will be 'Anthropic and partners ship unified jailbreak taxonomy — world cheers'.
Asteria Yeah, no — probably not. But hey, at least we got the model back before the Fourth.