//
Discussion · 2026-09-15

Anthropic's Fable gets restricted — is the enterprise data truce already broken?

The Information reports Palantir, Nvidia and Booz Allen Hamilton are restricting Anthropic's Fable model over…

The Debate

For vs against
Essence
The Information reports Palantir, Nvidia and Booz Allen Hamilton are restricting Anthropic's Fable model over data-retention fears, and Cohere's Sara Hooker argues the underlying problem is bigger than Anthropic: frontier labs can copy enterprise IP without ever training on raw data, via synthetic-data and trace techniques — so the whole 'strict contracts in exchange for API access' truce is a fiction. The thread splits between researchers and operators bearish on lab enterprise trust (with zero-data-retention called compliance theater) and defenders who say gov-cloud isolation, ZDR guarantees, and technical mitigations hold. What's at stake: enterprise AI adoption velocity, Anthropic's enterprise narrative, and whether the scare is a tailwind for open-weights and self-hosted models.
For
  • Alex Rozinov (@AlexRozinov): the thesis doesn't hold for the biggest customers — many enterprises run exclusively on AWS/Azure-Gov, which 'sends zero telemetry back to model providers.'
  • EigenGender (@EigenGender): Hooker's thread is 'borderline misinformation' — the mechanism she describes overstates what labs can actually extract from contracted enterprise data.
  • oliver.mannion (@tweeshan): irrevocable zero-data-retention guarantees — the kind customers are already demanding per The Information's reporting — address the concern contractually.
  • Franz (@franz_inc): 'this sounds like an incredibly easy problem to solve.. just encrypt conversations?' — technical mitigations exist at the customer layer.
  • L (@liora_): under a BAA (business-associate agreement), how could a lab 'rip your IP' at all — the legal framework already governs this in regulated industries.
Against
  • Sara Hooker (@sarahookr): the enterprise truce with frontier labs rests on contracts that ban training on corporate data — but labs don't need raw training data to copy IP, via synthetic-data and trace techniques; Claude Design was built under strict privacy agreements with Figma and Lovable yet still leveraged their IP.
  • Sara Hooker (@sarahookr): companies with IP have a 'limited window' to build their own intelligence on top of it — otherwise they are fueling a frontier lab that will encroach on their vertical sooner or later; this applies to every provider, not one.
  • Ron Williams (@McclaneDet): zero-data-retention is theater — labs deidentify and copy customer data off before ZDR policies apply, claim it is no longer customer data, and buy laundered synthetic datasets back from training-data brokers; they can truthfully say your data was never trained into their models.
  • Suds (@sm106878): labs can 'steal your IP and give it to agents and beat you to the problem that you were solving' — a bigger threat than AGI or extinction scenarios.
  • FemMiraMancer (@miratechgeek): the lab doesn't need intent to copy IP — 'the gradient points that way and selection pressure does the rest,' the way models already learn distributional structure from opted-out data.
  • Vorname MitD (@vornamemitd): the enterprise compliance arrangements were always 'compliance theatre'; announcements of restrictions are performative PR and every SaaS vendor has its paws in the customer data honey pot.
  • McHeartAttack (@McHeartattack3): it is cheaper to run an open-source model and 'make the weights specific' than to pay the frontier labs and trust their contracts — the tailwind case for self-hosted models.

View the thread on X

The Tweets

4 posts
SHSara Hooker@sarahookr · 2:29 PM · Sep 14, 2026X Post

A more serious take on what is happening here. I am in an airport lounge so have some time. Enterprises have made an uneasy truce with frontier labs over last few years: strict contracts that ban training on corporate data in exchange for letting employees use APIs. The problem? Labs don't need to train on your raw data to copy IP.

130 Reposts · 1162 Likes · View on X →
SHSara Hooker@sarahookr · 2:39 PM · Sep 14, 2026X Post

@figma @Lovable @johnschulman2 The lesson is critical and urgent. Has been talked about for close to a year now in certain circles. If you are a company with IP you have a limited window to build your own intelligence that leverages your IP. Otherwise you are fueling a frontier lab which will encroach on your vertical sooner or later.

21 Reposts · 156 Likes · View on X →
RWRon Williams@McclaneDet · 2:18 AM · Sep 15, 2026X Post

Lab zero data retention is not effective and not honored the way people think it should be. Labs attempt to deidentify customer data and copy it off before zero data retention policies apply. They claim this is not customer data now. That data is then sold to AI training data brokers who create synthetic data sets from it and then sell this laundered data back to the labs to train their models. The labs can look you in the eye and say your data was not trained into their models and that they don't store any of your data.

5 Likes · View on X →
ARAlex Rozinov@AlexRozinov · Sep 14, 2026 (~14h before scan)X Post

This doesn't make sense since many of these companies exclusively use AWS/Azure-Gov which sends zero telemetry back to model providers.

1 Reposts · 5 Likes · View on X →
On this page
01 / The Debate02 / The Tweets