@Claude is now your Slack coworker 🤖, the vuln report is dead 🔑, give your agent a DNS passport 🧑‍💻

Anthropic's Claude Tag adds Claude to Slack as a team member you grant channel access, tools, and codebases, then delegate work to by tagging @Claude. It remembers context from the channels it sits in and can plan multi-step tasks to run later. Anthropic says 65% of its product team's code now comes from its internal version, so this is already how the company operates.

The NSA lost access to Anthropic's Mythos 5 after the Trump administration's export controls forced the model's pullback, partway through the agency's own tests. In controlled red-team runs inside classified networks, Mythos identified security flaws fast, though officials pushed back on a senator's claim that it broke into those systems. Defense Secretary Hegseth has labelled Anthropic a supply-chain risk, and the company is now suing the government.

A supercomputer in Shenzhen called LineShine has been declared the world's fastest, putting China at the top of the rankings for the first time since 2017. The detail that matters for builders is how it got there: it uses only standard microprocessors, not the special-purpose GPUs that dominate AI compute today. It is a pointed reminder that the GPU is not the only road to frontier-scale performance.

Mistral's OCR 4 is a small document model that returns bounding boxes, typed-block classification, and inline confidence scores alongside the extracted text. It supports 170 languages, runs in one container for fully self-hosted deployments, and posts a 72% average win rate over rival OCR systems plus the top OlmOCRBench score. For anyone building RAG or enterprise search, that structured output drives source-grounded citations, redactions, and human review.

Stripe Directory, now in public preview, is a searchable index of businesses and services across its network that you or an agent can query from the CLI. Results come back as structured data: provider slugs, app listings, and machine-payment endpoints on mpp.dev that an agent can call and pay per request. It is an early piece of the plumbing for agents that discover and transact with services on their own.

The Linux Foundation said it intends to create the Agent Name Service, an open standard that gives AI agents verifiable identities tied to the domain name system. An operator proves control of a domain through ACME, the protocol behind Let's Encrypt, and gets issued agent certificates, with every status change written to an append-only log. For teams building multi-agent systems, it is a trust layer to evaluate before rolling your own.

A Stanford study tracked 3.4 million people submitting 4 million applications and found that AI screening tools, used by 90% of US employers and dominated by a few vendors, reject candidates along racial lines. Because one vendor's model influences many employers, a rejected applicant can be shut out everywhere they apply at once. It is the first large-scale look inside the black box that now gates entry-level hiring.

The SGLang team details how DeepSeek-V4 serving on NVIDIA's GB300 went from roughly 2,200 to 11,200 tokens per second per GPU at the same user-visible speed, a 5x gain since April's day-zero stack. The wins came from kernel and runtime work: MHC fusion, KV Compression V2, W4A4 MegaMoE, and better eviction and admission behaviour. The new curves hold throughput much deeper into the high-interactivity range that real deployments care about.

Meta's engineers explain how they shrank a battery narrow enough for smart-glasses temple arms, down to 7mm steel-can cells narrower than anything before. Traditional pouch cells waste volume in their folds and struggle to deliver peak power when the camera and an AI workload run at once, so Meta moved to rigid steel cans with stacked electrodes. The real constraint shaping wearable AI turns out to be peak power, not capacity.

Hugging Face describes how it took huggingface_hub, the Python client under transformers and datasets, from an update every four to six weeks to one every week. A single GitHub Actions workflow now handles version bumps, downstream CI triage, and AI-drafted release notes, using open tools and open-weights models with one human gate where judgment matters. Nothing in it needs a vendor contract, so other maintainers can lift the pipeline directly.

A new project posted to Hacker News, Neural Particle Automata, generalises Neural Cellular Automata from fixed pixel grids to free-moving particles, each carrying a position and internal state updated by one shared learned rule. To stop local interactions scaling quadratically, it swaps grid perception for differentiable smoothed-particle-hydrodynamics operators backed by memory-efficient CUDA kernels. It keeps the robustness and regeneration that make NCA interesting while handling morphogenesis, point-cloud classification, and texture synthesis.

Armin Ronacher names the pattern taking over agentic engineering: a harness-level loop outside the agent's own tool-calling loop that keeps a task alive past the point where the model would stop. Work goes into a queue, a machine attempts it, and the harness decides whether to continue, restart with fresh context, or hand off. He admits he has not made it work for code he deeply cares about.

Euclid's case is that the durable asset in vertical AI is not the model but the exception queue: the cases that do not fit the template, where a human overrides based on context that lives in no ERP or CRM. That reasoning normally evaporates the moment the case is cleared. Whoever captures it, the way Tennr owns referrals or Abridge owns clinical notes, captures the value, because agents already handle the easy work.

Go security veteran Filippo Valsorda says the social contract around vulnerability disclosure has quietly broken. Maintainers owed researchers fast acknowledgement and credit because confidential insight was scarce, but now any maintainer or attacker can run an LLM that turns up the same bugs. The bottleneck has moved from finding vulnerabilities to triaging which ones are real, so his advice is to stop treating reports as sacred and invest in triage and remediation.

InfoWorld makes the case that AI endpoints which tolerate fuzzy intent are reawakening service-oriented architecture, the discoverable-services dream that SOAP, WSDL, and the enterprise service bus buried under XML complexity in the 2000s. Where SOA 1.0 needed rigid contracts machines could not flex around, models can now mediate intent between services. If it holds, builders finally get the flexible, discoverable integration that SOA promised and never delivered.

Semafor's Liz Hoffman argues AI will commoditise plenty of knowledge work but not the kind Wall Street sells, because top-tier M&A advice is a Veblen good whose demand rises with its price, like a Hermes bag. Banks have never competed on price, since executives spend shareholders' money and the high fee is itself the signal. So even as AI automates the grunt work, expect fees at the top to hold.

Baidu open-sourced Unlimited-OCR, a model built to push past DeepSeek-OCR on what it calls one-shot long-horizon parsing, turning long documents into structured text in a single pass. The weights are on Hugging Face and ModelScope, and it runs through standard transformers on an NVIDIA GPU with a short dependency list. For anyone paying per page for document OCR, it is a free, self-hostable alternative worth benchmarking against the closed APIs.

Clips is an open-source, self-hostable alternative to Loom and Jam where every recording is built for agents. Paste a Clips link into a coding agent and it gets the transcript, summaries, and timestamped frames even when its model cannot ingest raw video, and a Jam-style capture hands over console errors and failed network requests alongside the clip. You scaffold one with a single npx command and own the player code.

OpenAI's new Codex Security plugin scans a chosen folder of code, either through a guided desktop flow or a single command in the Codex CLI. You install Codex, add the plugin, point it at a project, and it opens a chat with the scan ready to run. It folds vulnerability review into the same agent that writes the code, not a separate tool you must remember to run.

Treedocs is a Swift CLI that generates a one-line description for every file and folder in a repo, rendering current entries green and stale ones red so drift stands out at a glance. A git pre-commit hook can warn or fail the build when the docs no longer match the filesystem. The point is to give new teammates and coding agents concise structure, so they spend fewer tokens wandering the repo.