Pulse (last 3d) · Research 71 · Agents 69 · Products 58 · Industry 35 · Models 24 · Legal 21 · Infra 20 · Enterprise 19 · Media 17 · Open Source 14

Trending · Anthropic 18 · Claude Code 15 · OpenAI 15 · Claude 13 · Codex 11 · Google 10 · ChatGPT 9 · Cursor 9 · The Verge 7 · Slack 6 · Apple 5 · Gemini 3


The weak point this week is the plumbing around the models, not the models: Together.ai and OpenRouter have both documented prompt leakage and broken retention promises at hosted inference providers. The tools gaining ground all try to prove the same things: where the data went, what an agent changed, and what an agent was allowed to touch. Any PHI path through a third-party endpoint now warrants the breach assumptions applied to any data processor, and coding agents near regulated pipelines need a behavioral regression gate before they get more autonomy. together.ai | openrouter.ai

  • Together.ai disclosed prompt leakage between customer requests in June and July 2026. together.ai
  • OpenRouter's incident report asks whether inference providers actually honor their retention commitments. openrouter.ai
  • AIR records a service's behavior and turns any agent refactor into a PASS or a precise diff. producthunt.com
  • Zvi Mowshowitz has published his breakdown of the GPT-6.2 Astra and Sol system card. thezvi.substack.com
  • Gemini 4's new xhigh reasoning level brings a large jump in both quality and cost. simonwillison.net
  • A 26.4B MoE reportedly runs at 613 tokens per second on a single RTX 5090. x.com
  • Daraxonrasib, an oral KRAS inhibitor, roughly doubled survival in metastatic pancreatic cancer. theguardian.com
  • GPT-6.2 Astra and Sol: The System Card - Zvi's analysis of the safety disclosures for OpenAI's two new variants. thezvi.substack.com
  • Daraxonrasib for pancreatic cancer - KRAS-targeted pill roughly doubles survival in metastatic disease, hailed as the biggest advance in decades. theguardian.com
  • The most misunderstood graph in AI just got a major update - MIT Technology Review on a revision to a widely cited AI progress chart. technologyreview.com
  • AI Supply Chain Security: What Canary Tokens Found - Dragos reports what planted canary tokens exposed across the AI supply chain. dragos.com
  • Signal restricts mobile screenshots - Signal now calls AI surveillance an existential threat and limits on-device capture. techcrunch.com
  • Exa Websets V2 - Builds web datasets with 20% more matches, 4x the speed and half the cost, at $5 per 1,000 searches (vendor claim). producthunt.com
  • LLMs can read and write executables from a hex dump - Demonstrates byte-level machine code reading and writing by LLMs. sgiannopoulos.github.io
  • Can Anthropic afford to put safety first? - The Verge examines whether Anthropic's safety positioning survives its IPO and competitive pressure. theverge.com
  • Pandex - An MCP server serving official, versioned EU legal texts to agents, updated nightly. pandex.io
  • UN fails to agree on killer-robot rules - Talks on lethal autonomous weapons ended without agreement. theverge.com
  • Merrin - Personal AI that runs on-device and works offline, with end-to-end encrypted phone and desktop sync. producthunt.com
  • Dawnroom - Local image studio with 34 open models, an OpenAI-compatible localhost API and an MCP server. producthunt.com
  • GuideEye - A chest-mounted phone assistant that calls out hazards for blind and low-vision users, released early for feedback. producthunt.com
  • Samiksa - Turns papers into question-driven documents with sourced claims and sandboxes for reruns. producthunt.com
  • Omnilogue 2.0 - Local searchable library of sessions from Claude Code, Codex, Cursor and Copilot. producthunt.com
  • tinytldr - MIT-licensed templates that Claude or ChatGPT fill over MCP and export as images or video. producthunt.com
  • WASD Engine - Open-source game-building toolkit tuned for coding agents, with multiplayer hosting. producthunt.com
  • AI agents in your text messages - TechCrunch roundup of consumer agents that operate over SMS. techcrunch.com
  • Selling an AI version of your face - A microdrama startup pays people to license AI versions of their faces. techcrunch.com

Provider incidents, self-hosted routing and dataset audits all address one question: where regulated data actually travels.

Prompts crossed tenant boundaries at a major inference provider Together.ai disclosed prompt leakage between customer requests in June and July 2026, and OpenRouter published a broader account of data processing incidents across providers, including whether retention commitments are kept. A zero-retention clause is a contract term, not a technical control, so any PHI sent to a multi-tenant endpoint carries the same risk as data handed to an external processor. together.ai | openrouter.ai

Self-hosted gateways now catch silent model swaps and route to your own GPUs first AI Gateway puts self-hosted models and vendors like OpenAI, Groq and Together behind one OpenAI-compatible endpoint, sending traffic to your own GPUs before falling back to vendors. Wallras sits in the same position as a single container that detects drift and silent model swaps and keeps a tamper-evident log, which is the audit trail a validated clinical pipeline needs. producthunt.com | producthunt.com

Training-data provenance checks are becoming a clause-level product The Provenance Checker reads a list of datasets, never the data itself, and flags origin, license terms, personal and special-category data, who labelled it, version and declared train/test overlap, each with the clause behind the finding. Special-category flags map directly onto health data under GDPR, so this is the documentation regulators will ask for. producthunt.com

The new tooling for coding agents focuses on proving what changed and limiting what agents can reach.

AIR turns current service behavior into the test suite AIR records HTTP responses, database writes, messages, outbound calls, ordering and failures, then air protect returns PASS or an expected-versus-actual diff after any agent change, with approvals one scenario at a time. It works over MCP and in CI under Apache-2.0, which makes agent rewrites of legacy pipelines with no tests auditable. producthunt.com

Path-level access and anti-tamper gates for agents arrive in open source gitroots limits each developer or agent to specific repository paths, routes their changes through approvals and keeps a searchable audit trail. Aveyx targets agents that fake tests or leak secrets with a blast-radius analyzer, an anti-tamper diff gate and a Git rollback vault, fully offline under MIT. producthunt.com | producthunt.com

Google's Stitch Loop closes the loop from production metrics to pull request Stitch Loop reads a GitHub repo plus live logs, user behavior and feedback, builds fixes in a cloud sandbox and opens PRs, with built-in loops for security, latency and accessibility. Nothing merges without human review, and it keeps watching the target metric after launch. It is waitlist-only, with priority for AI Ultra users. producthunt.com

Agent infrastructure is settling into dedicated VMs, durable runtimes and team-wide memory.

Agent VMs now boot in a second and cost $0.04 per vCPU-hour Groundbox gives each agent a Linux VM that starts in about a second and can pause and fork with memory intact, at $0.04 per vCPU-hour with memory included. The same key runs GitHub Actions at half GitHub's price (vendor claim), so sandboxes and CI end up on one bill. producthunt.com

Bots with their own computer now run on existing Claude Code and Codex subscriptions Fenna Crew, under AGPL, gives scheduled bots their own browser and office editors on your machine, with human takeover and approval before risky actions. aiteam.space runs Claude Code sessions as "teammates" with their own computers, managed through an org chart. Both reuse subscriptions you already pay for. producthunt.com | producthunt.com

Year-long durable functions and branchable Postgres are now part of the agent platform Volcano bundles year-long durable functions, human-in-the-loop approvals, resumable sandboxes, distributed locks and branchable Postgres, deployable from Claude Code, Codex or Cursor. Long-running agent workflows are moving from custom orchestration to off-the-shelf primitives. producthunt.com

Shared agent memory now shows which agent wrote each entry and how often it was read Waku.One gives a team's Claude Code, Codex and Cursor agents one shared memory and hand-off board, with every memory attributed and its read counts visible. Memanto Cloud offers hosted memory for 34 agent clients with no Docker or API key. Shared memory becomes the next place for context to leak if it goes unaudited. producthunt.com | producthunt.com

Cost per answer is diverging: frontier reasoning tiers climb while strong open models fit on one consumer GPU.

Gemini 4's xhigh tier buys quality at a steep premium llm-gemini 0.38.1 adds the xhigh reasoning level for Gemini 4 Hydrogen and Neon, and Simon Willison's comparison shows a large jump in both quality and cost. That makes reasoning depth a per-task budget decision, not a model choice. simonwillison.net

A 26B looped-transformer MoE runs at 613 tokens per second on one RTX 5090 Basalt 0.5.5 reportedly serves Thinking Machines' Inkling 2 Mini, a 26.4B looped-transformer MoE, at up to 613 tokens per second on a single RTX 5090 using NVFP4 (single source). That throughput on one workstation card puts on-premises inference over sensitive cohorts within reach of a single desk. x.com

Frontier teachers now repair a student's reasoning at the token where it breaks In a new method, a frontier model finds where a student's reasoning fails, substitutes better tokens and trains on the correction, benchmarked with seven frontier teachers on Qwen2.5-7B (preprint, unreplicated). If it holds, small in-house models can inherit frontier reasoning more cheaply than with full distillation. x.com

10 items: bare links, memes, a hiring post, two Polymarket quote reposts, a career thread and pieces with no stated claim.