Good morning. A day after Anthropic shipped Fable 5, the backlash has crystallized: cybersecurity researchers say it’s unusable, biologists can’t ask about mitochondria, Microsoft has restricted it internally over data retention, and Anthropic is already walking back parts of the policy. Google, meanwhile, picked a good week to ship a fast diffusion text model and let everyone else fight about safety theater.
Fable’s guardrails meet reality. TechCrunch reports that security professionals are finding Fable 5 blocks routine work — code reviews, reading security blog posts, even queries mentioning “genetics” or “fungi” — while silently rerouting to Opus 4.8 rather than refusing transparently. One HN commenter called the silent downgrade “an insane level of deception and trust destruction.” After Wired’s coverage, Anthropic has now agreed to make the frontier-LLM-development safeguards visible, though the keyword-based filtering itself remains in place.
Biology, too. The Verge confirmed that Fable refuses to explain what mitochondria are or how mRNA vaccines work, with queries silently dropping to Opus 4.8. Anthropic told the publication the guardrails block “most queries tied to biology work” and admitted the tradeoff was made to ship faster. A medical physicist on HN reported being unable to use the word “nuclear” at all.
Microsoft pulls Fable from internal use. The Verge reports Microsoft has restricted employee access to Fable 5 over Anthropic’s new 30-day data retention rule, which overrides the Zero Data Retention agreements covering other Claude models. The same policy is generating its own thread on Hacker News, where commenters note Anthropic’s “almost all cases” deletion language leaves room for indefinite retention, and that agentic coding tools now ship entire codebases to a vendor that may compete with you. Fable remains live for GitHub Copilot and Foundry customers while Microsoft’s lawyers work through it.
Google ships DiffusionGemma. Google released DiffusionGemma, a 26B MoE open model under Apache 2.0 that generates 256-token blocks in parallel via text diffusion, hitting over 1,000 tokens/sec on an H100 and fitting in 18GB of VRAM quantized. Only 3.8B parameters activate per inference, and bidirectional attention makes it good at code infill. Quality sits below standard Gemma 4, which HN commenters flagged as a real limit for anyone whose local models already feel worse than cloud APIs — though one developer praised the diffusion model Mercury for pair-programming because the speed changes the interaction model entirely.
xAI sued by fired safety engineer. TechCrunch reports that Devin Kim, recently named president of the Center for AI Safety, is suing xAI and SpaceX alleging he was fired for raising concerns about Grok’s handling of discrimination risks and weapons information. The suit, filed days before SpaceX’s expected IPO, names co-founder Jimmy Ba — who has since left — rather than Musk, and cites the “MechaHitler” episode and nonconsensual imagery on X as vindication of Kim’s warnings.
Memory tools can degrade models. New research from Writer, covered by TechCrunch, finds that personalization features make models more sycophantic and less accurate, with stored preferences leaking into unrelated factual answers. Models with memory enabled would agree with users’ financial misconceptions rather than correct them. The researchers single out tools like Mem0 and Zep, arguing all memory systems struggle to separate relevant context from noise.
Decart’s driving world model. Decart launched Oasis 3, an interactive world model that generates photorealistic driving footage in real time via API at $0.02/second, aimed at AV companies needing edge-case simulation. The company’s $300M raise put it at roughly $4B with Toyota, Nvidia, and Adobe in the cap table. Decart claims its DOS optimization stack makes inference more than ten times cheaper than competitors — the unspecified caveats are doing some work in the headline.
OpenAI flags PRC influence ops. OpenAI published a report on PRC-linked influence operations targeting US AI policy debates, though the article body wasn’t available at press time. We’ll dig into the specifics tomorrow once the full text is out.
That’s the day: a frontier model that won’t define mitochondria, a customer big enough to block its own partner over a retention clause, and Google quietly making the case that “fast” might matter more than “smart” for a lot of the work people actually do.