Anthropic says its silent safety filter hits 0.03% of developers. The thread found false positives on base64

The problem isn't the refusal — it's the silence

The post by Jon Ready makes a narrow, technical argument about Anthropic's new Fable 5 safeguards. The objection isn't to visible refusals, which a developer can see and route around. It's that the model can be silently degraded — through prompt modification, steering vectors, or parameter-efficient fine-tuning — on requests Anthropic reads as "frontier LLM development." When that happens there's no message. So a developer staring at a bad answer can't distinguish three very different situations: the model is confused, the problem is genuinely hard, or a policy classifier has quietly throttled the response.

Ready disputes Anthropic's framing that this affects only 0.03% of developers, arguing the line between "frontier AI research" and ordinary business code is blurry enough that the restriction's blast radius is far wider than a rounding error. And he reads the competing-model clause less as safety than as a moat: invisible restrictions that happen to protect Anthropic's commercial position.

The base64 problem

The thread didn't argue the theory. It produced receipts.

"I tried today and it gave cybersecurity error on base64 implementation. It is so nerfed…." — a developer reporting a concrete false positive

"I am experimenting with Fable on a completely benign personal application and I am constantly hitting the 'cybersecurity and biology topics' guardrail" — another hitting the same wall

The most pointed report came from someone using Claude exactly as Anthropic recommends — to audit its own output:

"I spend a lot of time telling Opus 4.8 to search for security bugs in the code it wrote … Fable wont let me fix the security issues that Opus 4.8 created." — Mike Cardwell, an engineer working on security tooling

Even a self-described Anthropic loyalist drew a line:

"Silent sandbagging is not acceptable, especially given they've shown with this release their safety filters have HUGE amounts of false positives." — extr, a longtime Anthropic booster

"Now you're having to prove your innocence"

The structural worry is about burden of proof. If the model can degrade without telling you, every disappointing session becomes unfalsifiable on the vendor's side.

"this is their get out of jail free card … they can just say, oh well, you must have been doing something that triggered its distillation prevention technique … now you're having to prove your innocence instead of them having to prove you're guilty." — the founder of Dollhouse Research

One analogy crystallized why this feels different from ordinary content moderation:

"Imagine if Github said 'if we detect you're building a competitor to Github, we will silently degrade the results of your CI actions so that tests sometimes randomly fail'." — a commenter who works in applied ML

The counter-read: it falls back, and this was always true

Not everyone bought the alarm. One defender noted Fable degrades to Opus 4.8 rather than going dead — a weaker model, not no model. Others argued silent restriction is simply the standing condition of any corporate LLM, and that a competent engineer should be able to judge whether a problem is solvable using their own expertise rather than trusting the model's effort level. The strongest counter-fact, though, cuts toward Ready, not away: if benign base64 and personal apps are tripping the wire, "0.03%" is doing a lot of work. The numbers a vendor reports about its own filter are the one benchmark the thread couldn't independently check.