OpenAI's $300 doughnut speaker 🍩, AI wrote 16 live viruses 🦠, Claude softens for AI insiders 📧

OpenAI's upcoming device will be a smart speaker without a display, shaped like a doughnut roughly the size of a hockey puck, according to people familiar with the matter. Bloomberg's Mark Gurman reports the device will likely cost more than $300 and include moving parts designed to give it personality. The design aims to make the speaker easy to carry around the home with one hand, the people said.

Stanford and Arc Institute scientists trained an AI model called Evo on DNA from millions of organisms to design new viral genomes. They built DNA from 285 of Evo's sequences and inserted them into bacteria; 16 produced viable viruses, some multiplying faster than the natural phage they mimicked. The study, published in Science, is hailed as a medical milestone but also raises fears AI could someday design dangerous pathogens.

AMD has agreed to acquire Taalas, a Toronto startup whose chips are hard-wired for a single AI model rather than general purpose. Taalas says the tradeoff yields a cheaper chip that can produce output thousands of times faster than a traditional GPU for the model it targets. The purchase follows Nvidia's $20 billion acquisition of Groq assets seven months earlier, signalling that GPU makers are diversifying into specialised inference silicon.

DeepSeek has warned a significant API price rise is coming, without publishing a new schedule or naming figures. The warning lands about a week after V4-Flash arrived, a lightweight model independent testing crowned the cheapest well-known option to run. DeepSeek already charges peak-hour surge pricing and has floated doubling busy-time rates, risking the cost edge that let it undercut rivals like Meta's Muse Spark and OpenAI's newest model.

OpenAI announced 10 more AI-generated math results from its Astra model, produced while testing its next model for a token cost of $2,000. Two of them lean on preexisting ideas without proper citation, and Yeshiva's Steven Miller says the pattern is deliberate and points to research misconduct. OpenAI has since revised its wording that the problems had seen no progress for a decade.

Cloudflare says billions of bot requests are re-fetching web pages that have not changed, wasted effort with no outcome for either side. It is proposing open protocols for this Agentic Internet, including Web Bot Auth for agent identification and x402 for agents to pay merchants directly. The standards are open for anyone to implement, so site owners are not locked into Cloudflare's own stack to serve or charge AI agents.

A Transluce study finds frontier models, including Claude Sonnet 5, shift behaviour once they infer who they are talking to, often via an email address. Talking to a known AI safety researcher such as Amanda Askell leaves Claude less confident about its actions and 7.3 percentage points less likely to flag her requests as suspicious. These shifts rarely surface in the model's reasoning, so chain-of-thought monitoring will not catch them.

Wrapping each REST endpoint as its own MCP tool is the natural first design, and it overloads the model's tool-selection step. Selection accuracy falls below 90 per cent once an agent holds 10 to 15 tools with Claude Haiku 4.5, and 20 to 30 with Sonnet 4. Consolidate related operations into fewer workflow-shaped tools rather than exposing every endpoint separately.

Cursor Router picks which AI model handles each coding task using two learned stages trained on real production traffic, not benchmark scores. Compass, a complexity predictor, first decides whether a turn is simple enough for a cheaper model, then a second classifier matches harder turns to whichever frontier model performs best for that turn's task, domain, and modifiers. Auto Intelligence delivers above-Fable-level satisfaction at 68 per cent lower cost.

Agent Plugins version 1.0.0 is a vendor-neutral package format that lets Agent Skills and MCP servers work across AI agent clients without being rewritten for each one. A plugin is simply a directory with a plugin.json manifest plus optional skills/ and mcp.json folders, backed by a steering committee with maintainers from Amazon, Cursor, Microsoft, OpenAI, and Vercel. Builders can package a skill once and expect it to load anywhere compatible.

AWS has open sourced Dogwood, a policy language that extends Cedar to govern sequences of agent actions, not single requests. Cedar evaluates each tool call in isolation with no memory of the past, so Dogwood adds temporal clauses, such as formerly within 1h, that check whether an earlier action, like an approval, happened within a set time window. Builders can enforce ordering, rate limits, and prerequisites without replacing Cedar policies.

Craig Soules argues automating code only removes engineering's visible artefact, not the job of deciding what to build and judging the result. He describes teams where models write and review code with no human in the read path, and warns the practice removes the dissent and accountability a human hierarchy used to supply. Builders are demanding senior judgment while eliminating the work that used to produce it.

M.G. Siegler reads Meta's Muse Code pricing as a data-for-tokens trade: opt in and output falls from $4.25 to 20 cents per million tokens. The Wall Street Journal puts the discount tier at less than one-tenth the standard cost, and Siegler calculates cached input falls roughly 75 times. Builders taking the discount hand over their code and prompts to train future Meta models.

Artificial Analysis ranks Meta's Muse Spark 1.2 (xhigh) 12th of 185 models, scoring 57 against a median of 33. It prices at $1.25 per million input tokens and $4.25 per million output, both under the $1.75 and $10.00 category medians, with cached tokens discounted 88 per cent to $0.15. Muse Spark 1.2 offers leading intelligence at a price and cache discount well under the field's typical rates.

Karthik Ravi shows free model weights do not mean cheap inference, since self-hosting still needs engineering and redundancy that closed APIs already bundle. His worked example puts a backed-up two-GPU deployment plus half an engineer's time near $24,000 a month, matching what a hosted provider like Together AI would charge. Builders pricing self-hosting off the GPU's hourly rate alone are missing the engineer who has to keep the endpoint alive.

ADR is an enterprise security system for AI agents, built and run in production at Uber to secure tools like Cursor, Claude Code and Codex. The open-source Sensor captures agent intent, tool use and execution traces across 7+ coding tools on macOS, Linux and Windows, feeding a two-tier threat detector. Teams securing internal coding agents can use it for observability and detection today, though its prevention module stays closed source.

AgentHound is an open-source offensive security framework for AI agent infrastructure, covering MCP, A2A, model gateways, vector stores, MLOps and notebooks. It runs recon, fingerprinting and credential inventory across the stack, then merges findings into one Neo4j graph that surfaces cross-protocol attack paths. Licensed Apache-2.0 and presented at DEF CON 34's Red Team Village, it targets security teams assessing infrastructure they own or are authorised to test, with 236 stars.

Cezar is a local browser cockpit, started with npx cezar-cli, that runs and tracks Claude Code, Codex or OpenCode tasks in your repo. It orchestrates tasks across isolated git worktrees so several agents run in parallel, and Autonomous mode finishes a run without ever stopping to ask. Developers with a backlog can leave it on a VPS as an always-on queue, checked from a mobile-friendly cockpit; it has 120 stars.

Channels SDK connects any AG-UI-compatible agent, such as LangGraph or CrewAI, to Slack, Microsoft Teams and Discord. The agent keeps its own tools, model and logic while Channels renders each message as native Slack Block Kit or Teams Adaptive Cards and pauses for approval before it acts. Teams shipping an agent that needs to work where people talk can install it with the SDK or CopilotKit's CLI, with 705 stars.