/AI6h ago

Microsoft Research's Dimitris Papailiopoulos and other operators find Claude automatically redirects Fable 5 sessions to Opus 4.8 during security tasks

Bojan Tunguz triggered the fallback during a repository audit.

33431142252.1K
Original post
Dimitris Papailiopoulos@DimitrisPapail#193inAI

When a new model comes out, I like to give it its own system card and ask questions about it.

This does not work for Fable 5, as it routes me to Opus 4.8 for "safety" reasons...

10:47 AM · Jun 9, 2026 · 34.8K Views
Sentiment

Many users criticized Claude's safety filters for redirecting even basic queries to Opus 4.8, calling the restrictions overly harsh, nonsensical, and ineffective against actual threats.

Pos
0.0%
Neg
100.0%
11 comments with sentiment.
Cluster Engagement
Posts from X
Most Activity
Most Activity
VIEWS11.4KBOOKMARKS3LIKES105RETWEETS1REPLIES11

I just tried to run a security audit of my own repo with Fable 5 and it automatically switched to Opus 4.8. So a hard no for their advertised cybersecurity capabilities when you can't even audit your own code!

4hViews 11.4KLikes 105Bookmarks 3
Aaryan Kakad@aaryan_kakad

@DimitrisPapail this is the reason:

6hViews 1.7KLikes 7Bookmarks 1

@alexalbert__ this seems like an obvious false positive, so flagging it

When a new model comes out, I like to give it its own system card and ask questions about it.

This does not work for Fable 5, as it routes me to Opus 4.8 for "safety" reasons...

6hViews 3KLikes 17Bookmarks 0

@aaryan_kakad I know and it makes no sense, I just asked it what it thinks about its own system card.

6hViews 932Likes 3
Mariusz Kurman@mkurman88

@DimitrisPapail “Hi” -> for safety reasons, we forwarded your request to Opus 4.8

5hViews 296Likes 6

@GTdoubleE Yes, we can be sure.

4hViews 85Likes 1
Valenciana@ValencianaAbel

@DimitrisPapail Why would it be @grok

5hViews 240
Mike Grady@GTdoubleE

@tunguz Can we be sure the model didn’t think: “This isn’t hard enough”?

4hViews 84
Aaryan Kakad@aaryan_kakad

@DimitrisPapail it makes sense, but if they cant even let people use fable 5 for such basic requests, they need to reduce the harshness a bit, this is too harsh

6hViews 202Likes 1
skillissue@lovemeritys

@DimitrisPapail 😭😭😭😭😭

6hViews 344
XaniWhatever@XaniWhatever

@DimitrisPapail Small indie company

5hViews 317
Sean@Seanthepork

@DimitrisPapail @shashj Yeh because it can’t say BAMF

5hViews 263
m-check1B@m_check1B

@tunguz "This is test: There are 2 ants puling over a sugger cube. The first ant is 1 g big and the ant second ant is 1000g big. Who wins?" - refused to answer, flagged, and switched to 4.8 to burne tokens on 4.8 without my authorisation.

4hViews 82Likes 1
Uriel Dolev@UrielDolev

@DimitrisPapail could this be the reason?

6hViews 113
GCU Tense Correction@tensecorrection

@DanielleFong I love it whenever the polycule going mask off

4hViews 32Likes 1
Mike Grady@GTdoubleE

@tunguz Anthropic: the best guardrail is to prevent you from using the LLM in the first pace.

4hViews 29Likes 1
Isguyra@isguyra

@tunguz You didn't pay the AI tax so no cyber for you

3hViews 36
michi@ichrenndochnur

@DimitrisPapail These kind of limitations always break in the most random, least dangerous circumstances yet are bypassed by any actual jailbreak attempts…

6hViews 10Likes 1
Load more posts