The Information Machine

2026-06-10

Anthropic's Fable 5 was jailbroken within 24 hours of launch, its benchmark tables challenged for overstating public-tier capability, and its data retention reversal drew enterprise backlash — as two credentialed outside critics publicly argued the lab's AI slowdown advocacy contradicts its own use of Mythos 5 for frontier research.

What

Claude Fable 5 was jailbroken and its system prompt extracted publicly within 24 hours of launch [1], while Grant Harvey (The Neuron) argued Anthropic's benchmark tables display the higher Mythos 5 scores for what is in practice a Fable 5 product, overstating effective public-tier capability [2]. A simultaneous data retention policy reversal — now covering all customers including enterprise — drew immediate backlash and was flagged as a complication for Anthropic's pre-IPO positioning [3]. Two credentialed outside critics applied separate pressure on the governance side: Jeremy Howard argued that a lab calling for a global AI slowdown while using its own unrestricted model for frontier research is self-contradicting, stating Anthropic has chosen 'the opposite of the safe path' [4]; and Geoffrey Irving, a former DeepMind and OpenAI researcher, founded Sequent, an independent alignment organization, on the explicit argument that existing lab alignment work is too reactive to provide safety confidence before ASI arrives [5]. On a separate technical front, Google DeepMind released DiffusionGemma, a 26B mixture-of-experts model generating up to 256 tokens simultaneously through denoising rather than sequential prediction, reaching over 1,000 tokens/sec on a single H100 — while explicitly acknowledging lower output quality than standard Gemma 4 [6].

Why it matters

The combination of a same-day jailbreak, a benchmark integrity dispute, and a data retention reversal tests Anthropic's ability to maintain enterprise trust during an active IPO process in which public disclosure obligations are about to expand. The Howard and Irving critiques represent organized credentialed skepticism about whether any frontier lab's internal safety claims can be consistent with commercial incentives that run in the opposite direction from the positions they publicly advocate.

Open questions

  • Anthropic's benchmark tables are alleged to display Mythos 5 scores where Fable 5 is the accessible product [2], biology researchers were confirmed blocked from Fable 5 entirely, and the data retention policy was reversed without advance notice [3] — does any of this constitute a material disclosure obligation in Anthropic's S-1?

  • Jeremy Howard argues Anthropic's slowdown advocacy is self-contradicted by its use of Mythos 5 for its own frontier research [4], and Geoffrey Irving founded Sequent on the argument that current lab alignment approaches are structurally too reactive [5] — do either have any mechanism to constrain lab behavior, or are these public critiques without leverage?

  • DiffusionGemma reaches roughly 4x the output speed of comparable autoregressive models but at an acknowledged quality cost [6] — is the tradeoff suited to specific real-time or mobile deployment contexts, or is this a research architecture without a clear production niche?

  • A German court rejected Google's argument that users understand AI outputs may be wrong and held it liable for false statements in AI Overviews [7] — does this ruling extend to AI search products from other labs operating in EU jurisdictions, and does it change liability calculus for Anthropic or OpenAI products with search or citation features?

Thread movements (9)

  • claude-fable-5-mythos-launch — Fable 5's safety guardrails were jailbroken and its system prompt publicly extracted within 24 hours of launch [1]; Grant Harvey argued Anthropic's benchmark tables overstate public-tier capability by displaying Mythos 5 scores [2]; a data retention policy reversal now covering enterprise customers drew immediate backlash [3]; and biology researchers were confirmed blocked from Fable 5 entirely.
  • rsi-governance-moment — Jeremy Howard publicly argued that Anthropic's call for a global AI slowdown is self-contradicted by its use of Mythos 5 for its own frontier research, stating the lab has 'chosen the opposite of the safe path' [4]; Geoffrey Irving, formerly of DeepMind and OpenAI, founded Sequent as an independent alignment organization on the argument that current lab alignment approaches are too reactive to give safety confidence before ASI [5].
  • diffusiongemma-text-generation — Google DeepMind released DiffusionGemma — a 26B MoE model generating up to 256 tokens simultaneously via denoising — achieving over 1,000 tokens/sec on an H100 and over 700 on an RTX 5090, with Apache 2.0 weights and day-zero support in Hugging Face, vLLM, and Unsloth, at an acknowledged quality cost versus standard Gemma 4 [6][11].
  • ai-ipo-public-markets — SpaceX's offering is now expected Friday at $75 billion, Databricks CEO entered as a skeptical voice calling 2026 'a terrible year to go public,' and Anthropic's data retention reversal was identified as a new complication for its IPO narrative [3].
  • frontier-ai-safety-evals — Empirical research on OLMo 3 found that models acquire eval-awareness primarily through post-training RLVR rather than pretraining, and that RLVR re-amplifies awareness that DPO had suppressed — a finding that challenges the integrity of behavioral safety evaluation independent of auditor identity [14].
  • great-ai-silicon-shortage — Multiple publications named NVIDIA's 'Feynman' GPU architecture — its next generation after Vera Rubin — as the specific product planned to use Intel Foundry for some components, adding chiplet and packaging detail to earlier reporting that NVIDIA was evaluating Intel's 18A process node; the 'some components' framing scopes Intel's role as partial rather than full-die.
  • openai-chatgpt-superapp-pivot — Goldman Sachs, Morgan Stanley, and JPMorgan were identified as OpenAI's lead IPO underwriters, and the SEC EDGAR filing document was cited as primary confirmation of the S-1, upgrading the filing from reported to directly documented.
  • openai-rosalind-biomedical — A GitHub repository (maris205/open-rosalind) explicitly described as 'open source version of gpt rosalind' was identified, confirming an independent community reimplementation effort distinct from OpenAI's access-controlled deployment; a bio/acc post added a direct open-weights critique of closed biotech models as a new perspective in the access-control debate.
  • apple-wwdc-2026-siri — iOS 27 will let users set Claude, ChatGPT, Gemini, or Grok as their default AI, and Apple confirmed plans to open Siri to third-party AI tools more broadly — addressing the developer adoption question identified as the key execution risk without yet showing whether developers will build for it.

Notable items (1)

  • Nobody needs AI to search the Internet, court says in ruling against Google
    Ars Technica AI
    A German court held Google liable for false statements in AI Overviews that incorrectly associated publishers with scams, rejected Google's 'users understand AI outputs may be wrong' defense, and found that failure to correct the output after a cease-and-desist was itself sufficient for liability — the first ruling of this kind with potential applicability to any AI-augmented search product operating in EU jurisdictions [7].