Claude Fable 5 updates biology safeguards, reducing fallbacks

#1 𝕏

Claude Fable 5’s biology safeguards are being updated to reduce false positives. In testing, the update reduced biology-related fallbacks by about 85% across the referenced product surfaces, enabling assistance with a wider range of everyday health and educational questions. Requests considered dual-use—including virology, toxicology, and molecular design—will continue to fall back to Opus 5, so Fable is not yet usable for professional biology research and drug development.

#2 𝕏

Google DeepMind shared a link to its blog about Gemini Robotics 2 and whole-body intelligence for robots.

Also covered by: @Google DeepMind

#3 📝 OpenAI News

Responding to the next frontier of critical cyber capabilities - OpenAI outlines steps to address emerging critical cyber capabilities, emphasizing security assessments and responsible evaluation. The post describes engagement with external partners and measures to strengthen model resilience and oversight.

Also covered by: @OpenAI, @Sam Altman

#4 𝕏

Harrison Chase shared his perspective on the journey from early LangChain to managed agents, describing managed deepagents as one of the launches he had been more excited about in a while. He said managed agents could significantly improve how easy it is to run agents.

#5 𝕏

Hugging Face shared a broadcast about how AI agents reproduced ICML 2026 papers.

#6 𝕏

Boris Cherny said stacking model training, input probes, and an intent-checking classifier can bring indirect prompt injection to ~0 on unseen attacks—a result he did not expect a year ago. He also announced that auto mode would become the default in Claude Code the following week.

#7 ▶️

Total Recall: The Invisible Memory Layer for AI Agents

SyntaxGTM

#8 𝕏

Santiago described Open Tag, announced by CopilotKit, as an open-source version of Claude Tag for deploying agents in Slack and Microsoft Teams. He said it lets users swap harnesses and models and move their data.

#9 𝕏

Peter Yang announced improvements to petergyang/human-review, a 100% free project with 500+ GitHub stars. It now supports Markdown-style lists, links via Command-K, drag-and-drop images, and multi-page review and editing via Command-click.

Also covered by: @Peter Yang

#10 𝕏

v0 released its new Usage & Activity Dashboard for tracking daily credit usage, analyzing activity by member or project, and drilling into messages by date, model, and cost. It’s available under Settings → Usage & Activity.

#11 𝕏

Santiago explains that harness-level agent loops repeat four steps and require two essentials—memory and stop conditions—with four safeguards: a 10-iteration cap, timeout, stopping after three identical tool calls in a row, and a goal check. Oracle, which partnered with Santiago, published an article covering three levels of loop design and shared a notebook implementing all three.

#12 𝕏

claire vo demonstrated how she used Codex to turn something into an always-on toolbar app for managing her smart lightbulb from her Mac, in a post referencing a video.

#13 𝕏

Guillermo Rauch recommends server-rendering and streaming at least the data, even when rendering on the client, warning that waterfalls of `fetch()` calls from hooks impose too high a performance cost.

#14 𝕏

Julien Chaumond expressed excitement about AMD’s acquisition of Taalas. He described “the model is the computer” as an early approach that could make inference faster, cheaper, and more energy-efficient, and pointed to a possible future in which inference is ubiquitous.

#15 𝕏

Guillermo Rauch recapped feedback from an unidentified tech lead for an AI agent platform built on eve.dev at a company with more than 55,000 people: the team found @aisdk good but too low-level, off-the-shelf solutions and unspecified Enterprise products expensive and inflexible, and agent frameworks inadequate. Rauch praised the @evedev_ team and said it is hard to find an abstraction that is both easy and scales with sophistication.

#16 in

Anu Jagga Narang recapped her remarks at ITX that AI strategy depends on four company-specific factors: the customer problem, technology, cost constraints, and definition of success. She suggested questioning whose context a playbook reflects and emphasized that strategy is only as good as its implementation.

#17 𝕏

Shreyas Doshi commented that 4.8 felt right for him, while 5.0 removed the personality that made Claude distinctive. He characterized 5.0 as more confidently wrong than previous models and as lacking their openness and unusual sentence construction.

#18 𝕏

Thariq described automode as safer than other permission systems, including manual review. Quoting an official ClaudeDevs announcement, he noted that automode was being rolled out to everyone by default with no classifier overhead cost.

#19 ▶️

AI is getting a little out of control

AI Explained

#20 𝕏

Yann LeCun congratulated Demis and welcomed him to what LeCun described as the club of former AI executives turned chief scientists.