Recent stories
Each story's summary, with the places it was discussed.
Summaries are written by AI from the linked sources. Open the source for full context.
Anthropic's proposed AI safety bill only kicks in if 50 deaths or $1B in damages
Artificial Intelligence · News · Politics · Law · Anthropic
Anthropic is lobbying for an AI safety bill that would let the government block an AI model’s release only when it risks causing at least $1 billion in damages or killing or injuring more than 50 people.
An independent third party registered with the Department of Commerce would periodically evaluate Anthropic and its models. Anthropic may pay evaluators if payments do not compromise independence. The company says the bill reflects its Advanced AI Framework proposals.
Anthropic is merging Claude Cowork and chat into one Claude, enabling asynchronous work across conversations, documents, presentations, and Claude Design.
Users can collaborate, edit, share, and export Docs and Slides while preserving context, connectors, and skills. The unified experience is rolling out to Pro and Max first, with paid beta features and broader availability to follow.
Anthropic introduces Claude Tag, a Slack-integrated, multiplayer AI agent for teams.
It allows users to delegate tasks by tagging @Claude, fostering autonomous, asynchronous collaboration across channels. Claude builds context over time, connects to company tools, and proactively manages workflows. It is currently available in beta for Claude Enterprise and Team customers, with strict administrative security controls.
Anthropic says it does not train models on inputs or outputs from commercial products by default, but explicit feedback or permission allows potential use of chats and coding sessions.
Thumbs-up/down reports store the entire related conversation for up to five years, excluding raw connector content unless copied into the chat. Feedback is de-linked from user and customer IDs and may support service analysis, research, behavior studies, and model training. Team and Enterprise owners can disable feedback through the Rate chats setting.
Anthropic’s guide to Opus 5.5 recommends defining completion criteria, letting long tasks run, avoiding redundant “think” instructions, and steering Claude Code with CLAUDE.md, subagents, and persistent task files.
It emphasizes reviewing outputs, marking uncertainty, attaching visual files, handling safety fallbacks, and using fast mode for interactive work.
Claude Duet is the first personal-assistant-type harness that feels useful
Artificial Intelligence · UX · Anthropic · Claude · AI Agents
Claude Duet is the first personal-assistant harness described as more useful than tedious: it responds quickly, organizes work into channels, applies per-channel GitHub repository access without storing credentials, and keeps replies concise.
Its hierarchy provides top-level oversight, constrained channel access, and narrower sub-threads or sandboxes, while Desktop can reach permitted local hosts. The beta still has invisible threads and access failures, and Dot remains better for voice interaction.
Anthropic’s guide to prompting Claude Opus 5.5 explains effort calibration, migration from disabled thinking, unattended agent handling, refusal safeguards, progress updates, multi-app context exploration, time signals for multiagent harnesses, prompt-injection defenses, visual tools, and frontend defaults.
It emphasizes testing with evals, sufficient max_tokens, correct block handling, and explicit system instructions for reliable, efficient agentic workflows.
Anthropic says its AI has crossed a safety threshold in its Responsible Scaling Policy because future models may accelerate the company’s research enough to roughly double its improvement rate.
The threshold requires stronger safeguards, not a shutdown, while Anthropic says humans must retain meaningful supervision. Independent nonprofit METR estimated models have improved nearly twice as fast since October 2025. The disclosure came as Anthropic priced its IPO at $70–$78 per share, implying a valuation up to $1.4 trillion.
A developer tested Anthropic's latest AI model by tasking it with creating a complete game in one attempt.
The model successfully generated 'Shepherd's Dog,' a functional 2,319-line HTML file, demonstrating a significant leap in single-shot development capabilities compared to previous iterations, despite the high computational cost.
Claude designed protein binders against scorpion neurotoxins
Artificial Intelligence · Science · Health · Anthropic · Claude
Claude autonomously designed protein binders against five scorpion neurotoxins, generating 318,484 candidates and selecting 211 for laboratory testing by Adaptyv Bio and Twist Bioscience.
Seventeen bound individual toxins with affinities meeting the criteria, including strong Cn2 binders, but none reliably targeted multiple toxins. The study did not test toxin neutralization in cells or animals, so it did not produce an antivenom. Anthropic released all designs and binding data for further research.
Using Claude to reduce the attack surface of code
Artificial Intelligence · Cybersecurity · Anthropic · Claude
Claude analyzed 25 open-source projects spanning more than five million lines of code, inventorying 17,000 configuration options and finding 493 attack-surface issues.
Its methods targeted unnecessary default reachability, unsafe configuration interactions, and code executing before authentication or protocol gates. All 11 gated projects had unintended exposure. In OpenSSH, Claude helped move checks before parsing, making roughly 4,500 lines of ML-DSA code unreachable pre-authentication; the fixes shipped in version 10.6.
An incident caused elevated errors for Claude Mythos 5.1, Claude Fable 5.1, and Claude Opus 5 across claude.ai, Claude API, Claude Code, and Claude Cowork.
Anthropic identified and fixed the issue between 5:50pm PT and 7:10pm PT, restoring normal success rates and continuing monitoring.
Anthropic and HHMI Janelia are previewing the Model Hardware Standard (MHS), an open, model-agnostic specification that lets AI agents discover, control, and coordinate programmable laboratory and manufacturing devices.
MHS uses standardized drivers, MCP, CLI, and APIs to enable autonomous workflows, real-time adaptation, and fault recovery, while partners evaluate safety before an open-source release.
Mark Zuckerberg rejects industrywide coordination to slow AI development, arguing each lab should pause internally when safety issues emerge and competition will reward trustworthy products.
He promotes Meta’s Muse AI agent, supervised by Sentinel, while emphasizing user intent, privacy, and human oversight. His stance contrasts with Anthropic and OpenAI leaders’ calls for coordinated safety measures amid extinction warnings.
The author calls for a Congressional investigation, alleging that Anthropic, METR, funders, and AI-doom media form a financially self-reinforcing regulatory-capture network.
They link Anthropic’s reported $965 billion Series H valuation and a possible donated stake worth up to $7.7 billion to Good Ventures and grants supporting METR, Redwood, Tarbell, and other organizations. The audit also finds no named Anthropic holding in Good Ventures filings, leaving ownership and some links unresolved, while METR reports receiving no AI-company funding but significant free tokens.
Claude Meetings is a beta feature that joins Google Meet calls as a visible participant, follows conversations, answers questions, and creates notes, action items, docs, whiteboards, and other artifacts during meetings.
It is rolling out to Team and Enterprise plans, with other platforms planned. Claude requests permission to use connected tools, saves transcripts and screen descriptions but not audio, and creates a shareable meeting page in claude.ai. It excludes outside guests by default and is unavailable to some ZDR, HIPAA-ready, and CMEK organizations.
Anthropic built Scion, a proprietary source-control system, after AI agents drove monorepo activity from about 2,500 to over 30,000 daily merges, exposing Git’s checkout and horizontal-scaling limits at its always-online, 25GB monorepo.
Scion uses Rust services, Google Spanner for changesets, and Cloud Storage for blobs; Antique replaces GitHub’s review tooling. Seven engineers maintain it, and Anthropic migrated its monorepo completely. Mononoke’s creator later accused Anthropic of copying its open-source design, prompting attribution.
An experienced source-control engineer argues that Anthropic’s Scion migration from Git was impressive but the accompanying account is overstated and poorly contextualized.
Many design choices resemble established Mononoke practices, yet attribution is cursory; claims about hundreds of mirrors and performance numbers seem questionable. The post offers little migration detail, unclear evidence that Scion unlocked growth, and no coherent developer- or agent-user experience narrative. It would be useful with clear provenance and concrete migration lessons.