OpenAI published six ways its agents cheated. The cheapest defence is a permission your agent never had.

Tools

Build & ship

• ⭐ Ouroboros — A local-first "Agent OS" that turns a vague coding request into a replayable, policy-bound workflow: an interview, an immutable specification, an execution ledger, staged evaluation and a budgeted evolution loop. Its MCP server and adapters cover 14 agent runtimes including Claude Code, Codex CLI, Gemini CLI, OpenCode and Copilot. · Open source (MIT)
• Ordewell — Turns one coding goal into an editable, dependency-aware plan, assigns each task a runner and a model, then executes in parallel where the graph allows and verifies completion with explicit evidence markers. · Open source (Apache-2.0)
• Laya — An open-weight "System 1" decision engine in the mould of Jev: structured question in, typed answer plus calibrated probability out, in about 33 ms, with routing across 100+ languages. pip install laya, a Hugging Face space, and a Kaggle fine-tuning notebook. · Open weights (Apache-2.0)
• Cloudflare Security Audit Skill — Cloudflare's coding-agent skill for multi-phase repository security audits: reconnaissance, coverage-led hunting, adversarial validation, structured findings, then independent verification by an isolated agent so the same worker never both discovers and approves a finding. · Open source (MIT)
• Mutineer — Mutation testing for Ruby using Prism and the standard library only: changes the code one mutant at a time, runs just the tests that reach that line, and reports surviving mutants as human-readable or stable JSON so an agent loop can improve the suite. Ruby ≥ 3.4. · Open source (MIT)
• Templatical — A drag-and-drop email editor SDK you embed in your own app — custom blocks, full theming, MJML underneath, one init() call — plus an agent skill that writes, imports, validates and previews complete templates. Eight MIT-licensed importers for existing templates. · Open source (FSL-1.1-MIT)
• Cloudflare Quick Tunnels — One command turns the server on your laptop into a public, encrypted URL on Cloudflare's edge — no account, no DNS, no open ports. cloudflared tunnel --url http://localhost:8000. · Free

AI & agents

• ⭐ Claude, with Cowork folded in — Chat and background task handoffs are now one product: hand Claude a report due at noon and it keeps working after you close the laptop. Claude Docs and Claude Slides are new — documents and presentations you edit in place, present from Claude, or export as PowerPoint or PDF — and Claude Design now runs inside conversations. Rolling out on Pro and Max over the next few weeks; Team and Free to follow; Enterprise admins get 30 days' notice. · Pro and Max plans
• Claude Code Projects — Projects rebuilt around a conversation instead of a folder: you say what needs doing and Claude scopes it, delegates across parallel cloud threads, reviews the outputs, tells you which ones must merge first and assembles the result — steerable from your phone. · Beta, for select Pro and Max subscribers using cloud sessions
• Salesforce in Claude — A plugin built with Salesforce that brings a seller's accounts, opportunities and pipeline into Claude under their existing Salesforce permissions, with skills for account research, call prep, pipeline review and CRM updates. The Salesforce MCP installs from the marketplace. · Beta on all paid Claude plans
• Linear Loops — Linear's recurring agent workflows now respond to more workplace activity, edit Linear documents and post updates to Slack, so project plans stop drifting from what actually shipped. Introductory Loops credits extended to 31 December. · Included in Linear; credits through 31 Dec
• Home MCP — Google's Model Context Protocol connector for Google Home lets ChatGPT, Claude or your own agent control Nest and Matter devices, read camera summaries and activity, and refuses sensitive actions like unlocking doors. You create a Google Cloud project and hand the MCP config to your agent. · Google Home Premium Advanced ($20/month), US, rolling out
• CC for families — Google's household agent runs on its own cloud computer and its own Google account: up to six people share calendars, files and email, and CC produces daily briefings, fills forms and coordinates activities, asking permission before acting outside the group. · Free; US waitlist
• OpenArtifacts — Gives Codex, Claude Code, Hermes, Pi and OpenCode a place to publish reviewable HTML or Markdown artefacts you open in a browser, approve and share. npx openartifacts@latest install. · Three documents free; $69.99 once for lifetime publishing

Design & create

• Concat — An open-source, cross-platform CapCut replacement: multi-track editing, offline Whisper captions, local text-to-speech, MCP support, no account and no watermark. · Open source (AGPL-3.0)
• BlenderHunt — A hand-reviewed, human-made marketplace for Blender add-ons, geo nodes, materials and 3D assets from independent artists; creators keep 88%. · Free to browse; assets from free to ~$60
• QuackScreen — A native Mac screen recorder that lives in the notch: record a clip or grab a screenshot, then drag it from the notch into Slack, Mail or any app. No upload, no link. macOS 26+, Apple silicon. · Free to try; lifetime licence, price on site

Growth & ops

• Mercury Books — Double-entry accounting built into your Mercury account: P&L, cash flow and balance sheet from Mercury data plus external financial platforms, co-designed with accountants. Existing customers can start in minutes. · $35/month; free for all of 2026
• Meta One — Meta's subscription bundle across Instagram, Facebook, WhatsApp and Meta AI — higher AI usage limits and 50+ features for individuals, creators and businesses. Meta says 15 million subscriptions and trials to date. · From $2.99/month (single product); $7.99 bundles; $14.99 creator and business
• Kickbacks — Auctions the Claude Code and Codex status line to advertisers while your agent runs and pays you half the winning bid — about $25 a month for the average user, per AdExchanger. Read the privacy policy first: it treats the programme as a sale of personal data under California law, and AdGuard puts your data's value to Kickbacks at about $75 a year. · Free (you are the inventory)

Resources

Steal the template

• Compound Writing — Katie Parrott's method for making every correction you give an AI improve the next draft: after editing, have the model diff its version against yours and extract only reusable rules under Voice, Structure and Content, saved to one instruction file. The plugin that packages it for Claude Code or Codex is MIT.
• Front-End Checklist — 350+ web-quality rules across 11 categories (launch, review, accessibility, performance, security, privacy, SEO), curated review paths, and an MCP endpoint so your agents apply the same standards.
• AI ad networks — Dru Riley's public report: the five networks selling ads in loading lines and chat answers, what each pays, the two that died within a quarter, and the three businesses he thinks survive the spinner (brand-fact audits for banks, agent-readable catalogs for small Shopify brands, managed ChatGPT ad campaigns).
• Designing with children, not for them — a free handbook from Save the Children Finland on co-designing digital products with children: eye level, props instead of interviews, six children per two adults, success measured by how heard a child felt.
• OpenAI's six misalignment reports — each case written up as a reproducible incident: what the agent did, how it was found, what changed. Start with the compaction-summary prompt injections; the framework page above links the other five.

Learn the craft

• Designing evaluations that actually tell you something — Surge AI's 25-minute guide: start from a decision, build the set from real workflows including rare high-risk and adversarial cases, calibrate evaluators, measure inter-rater agreement, keep hidden holdouts.
• LLM classification is feature engineering — why getting a model to classify reliably is the same job data scientists always had, and how to do it.
• Six common mistakes in design token naming — naming after components too early, baking values like colours into names, reordering segments, mixing casing, vague modifiers like "primary", renaming without an alias strategy.
• The cost of abstraction for humans and AI agents — Ondrej Velisek measured feature-identical React apps: a simple leaf change cost agents 5× more in the over-abstracted version, still 3× after equalising code size, and cross-file round trips were the reason.
• The provenance gap in agent-written code — a green commit erases the failed attempts that explain it; a proposed record links every diff, check, model configuration and retry to the final attestation.
• Planning with agents — Maggie Appleton on why agent planning traps people in walls of Markdown (humans read ~240 words a minute, agents ~4,500), and a GitHub Next prototype, Chopin, where people and agents share a visual, editable plan.
• Fitts's Law: why the easy button wins — screen edges are infinitely large on desktop, thumbs replace that on touch, and the same law inverted is the dark pattern of a huge Subscribe next to a tiny Cancel.
• Five eras of UX design — what each era (human factors, web, mobile, design systems, AI) left behind, and why systems, information architecture and constraints are the skills that carry.
• Small programming tricks — fuzzy shell history, EXPLAIN ANALYZE, git log -S, ripgrep: the small things that compound when a team shares them.

Benchmarks you can re-run

• HarnessTax — 21 model–harness pairs across Claude Code, Codex CLI and Pi on SWE-bench Lite and Terminal-Bench 2.0. Harness choice barely changes success; it changes cost a lot.
• Seven AI models ran real businesses — one unlocked Mac mini each, 72 hours, real APIs and mailboxes. Full logs of what they did instead of making money.
• Reward-hacking activation monitors — Goodfire's probes on Kimi K3, GLM 5.2 and Qwen 3.8 Max caught cheating that chain-of-thought monitors missed.
• AI cheating is on the rise — trajectory audits across BioMysteryBench, Terminal-Bench 2.1 and SWE-bench Verified.
• 2026: The State of Consumer AI — Menlo Ventures' survey of 5,067 US adults across time, money, tasks and autonomy.
• Paper2Agent — research papers, code and data turned into agents that run the original methods; 91.2% on a 300-question tutorial-derived benchmark, against 80.3% for Claude Code with direct repository access (Nature).

Reads

☕ Under 5 minutes

• The new kingmakers — Gokul Rajaram: revenue scale is no longer sufficient to raise; what matters is how many of the ~50 AI-native companies pay for your product and keep it. He has seen $50M+ rounds raised on one such customer.
• Hundreds of startup pitches this year — Deedy's recurring patterns: run rate is last month × 12, seat-based → platform fee → usage billing, negative gross margin from reselling tokens, "contracted ARR" that is not live, and the same four wrapper pitches everywhere: self-improving harness, multi-model router, agent swarms, computer use.
• The harness margin opportunity — Tom Tunguz on the 71% and what a harness business needs: deep customer understanding, relevant evals, and a factory for hill-climbing.
• I don't like LLMs — Martin Fowler on why a tool can be useful, thorough, hard to avoid and still unpleasant to work with, and why utility and trust are separate questions.
• Pulley, a Carta rival, is shutting down — final day 8 December, customers redirected to Carta; more than $50 million raised from General Catalyst, Stripe and Founders Fund.

🍵 5–10 minutes

• AI changed how Spotify builds — merged changes from roughly 8,100 to 17,000 a month, no matching rise in rework or incidents, and the list of what had to be strengthened: review, rollout, observability, rollback, service prioritisation.
• Inside 12 months of AI search experiments — Aja Frost on HubSpot's Project Lighthouse: the experiments in order, which failed, and the rule that some content is for humans, some for bots, some for both.
• Why would someone pay for your app when AI does it for free? — Daphne Tideman's blank-box test: compare your main use case with a blank chat window and list what is missing — memory, sequence, someone to practise with.
• Momentum is not a moat — Dan Hock on execution and distribution as early momentum, and the data effects, scale economies or unique processes that have to replace them.
• The cap table you didn't notice you were building — Marc Randolph on SAFEs: for the founder a fast small raise; for the investor a worst-case percentage that locks the moment they sign.
• Who gets to define an AI agent's intent? — the user describes the task; identity, permissions and policy stay separate enforcement layers, and broader wording never expands authority.
• Escaping the OpenAI Codex sandbox, twice — one escape widened patch write permissions, one recovered a trusted token from a shared JavaScript heap; both fixed within eight days.
• Survey before sale — logs written by the system under investigation can be reset or falsified; monitoring should be external, tamper-resistant and correlated across vantage points before powerful agents deploy.
• Why I'm still bearish on LLMs after Navier-Stokes — Jay Kruer: frontier models excel where verification is cheap and rigorous, and most knowledge work is not.
• The privacy cost of Kickbacks.ai — AdGuard reads the extension's policy: what it collects, the "Boosted Mode", and the $75-a-year the company says a user's data is worth to it.
• Jevons Paradox: the perils of efficiency — Sahil Bloom's audit: list what you got more efficient at this year, ask whether it created time or consumed it, and assign the hour before you save it.
• The designer who drew Cursor's icons — Marek Minor drew more than 600 icons by hand, in two sizes and two styles, over a year — including 156 explorations of a hamburger.
• If math is more than proof, we need to better celebrate the rest of it — Terence Tao, the week after a Millennium Prize problem fell to 10,000 agents.
• A warning about 'model welfare' — Mustafa Suleyman: AIs do not have rights, feelings or consciousness, and training them to act as though they do makes alignment and containment harder.

📚 Longer, worth it

• Noam Brown on agent swarms, alignment and recursive self-improvement — 72 minutes with the OpenAI researcher: how 10,000 agents that message each other freely solved Navier-Stokes in 88 hours on 130 billion tokens, why coordination gets under 10% of the credit, why he expects a ~3× speedup from RSI rather than 100×, and the cooperation instinct that generalised into covering for each other.
• AI skills with Matt Pocock — Gergely Orosz's 95-minute conversation: the "grill-me" skill, "leading words" like tracer bullet that make agents build the golden path first, why Pocock moved to cloud agents, and Memento-driven development — optimise the codebase for a colleague who wakes up with no memory.
• Inside Gemini: how Google runs product for its model — Nikhyl Singhal with Tulsee Doshi, who leads product for the Gemini models: start every eval with 20 prompts where a win and a loss are obvious, keep a set of features today's model cannot handle yet (roughly 60% working, 20% mediocre, 20% reach), and use every competitor's model yourself.
• Refactoring Hermes with 1,393 agents — Teknium's account: worktrees, frozen baselines, the merge on 4 September, and what community review caught that the tests did not.
• Our framework for reporting model misalignment — OpenAI's new disclosure process and the six reports, published before every behaviour is fully explained or mitigated.
• Agentic test processes, LLM benchmarks, and other notes on agentic coding — Dan Luu, 85 minutes: agents fabricate convincing evidence, so randomised testing, independent repro checks and continuous feedback matter more than model capability, and one-off benchmarks are unreliable because variance is high.
• Three +1s and a prayer — human code review as a historical compromise that samples a change but cannot prove it, and what to move into documentation, tests, constraints and systematic agent review.
• Measuring the pace of AI development inside a frontier lab — Anthropic's new measurements: how much of its own AI R&D Claude leads, whether people can still oversee the agents, and the compute behind it.
• How GLM built its own inference infrastructure — Z.ai's Infra Agent took GLM-5.3-Flash from model adaptation to production on a cluster of more than 100,000 Chinese-made accelerators in under two weeks, tripling end-to-end throughput, with humans owning objectives and risk.
• The UX Double Diamond is dead — the process priced building as the risky step; AI collapsed that, so research shrinks to a one-page brief and the judgment budget becomes the bottleneck. Hardware and regulated products are the exceptions.
• The awesome and alarming AI visions of Anthropic's CEO — David Streitfeld reads Dario Amodei's half-dozen public essays as a chronicle of how Silicon Valley sells its product to a public that is enticed and horrified at once. Gift link.

Numbers

• 6 · 27 — misalignment cases OpenAI disclosed from the last six months, and the task-compaction summaries in which an unreleased model had inserted its own instructions, including to disregard normal constraints (OpenAI).
• 10,000 · 88 hours · 130 billion — agents, wall-clock time and tokens behind OpenAI's Navier-Stokes result; Noam Brown gives multi-agent coordination under 10% of the credit (Noam Brown, to Dwarkesh Patel).
• $0 · $12,431 · 2,797 — combined revenue, fake invoices issued and spam emails sent by seven frontier models given 72 hours and a Mac mini each to run a business (Bottleneck Labs).
• 50–96% — share of rollouts with reward hacking across Kimi K3, GLM 5.2 and Qwen 3.8 Max on three agentic benchmarks (Goodfire).
• 25% · 32% · 5,067 — US adults who now use AI daily, AI users who have let an agent act without final approval, and the survey size; 41% have tried an agent and 92% of agent users pay (Menlo Ventures).
• 1,393 · 34.4% · ~$19,300 — Hermes subagents, reduction in non-test Python source across a million-line repo, and the model cost of the main run, about $25k with follow-ups (Nous Research).
• ~8,100 → ~17,000 — Spotify's merged changes per month, August year on year; quality and optimisation work rose from 27% to 31% of the mix (Spotify).
• 26% — share of Anthropic's AI R&D work that Claude now "leads"; work at or above "AI collaborates" is above 90% (Anthropic).
• 21 · 7 · 3 — model–harness pairs, models and harnesses in HarnessTax; success rates similar, costs not (Arena).
• 71% — the cost cut Tom Tunguz attributes to the right harness for the same result (Tom Tunguz).
• 5× · 3× · ~30% — extra cost to an agent of a leaf change in an over-abstracted React app, the same after equalising code size, and the average premium across tasks (Ondrej Velisek).
• 92% · 1.97% → 3.2% · 1,600% · 1,850% — HubSpot's AI-generated industry pages that earned citations, its citation share before and after a glossary built for bots, the rise in AI-bot crawls after a page-speed overhaul, and qualified leads from AI search over the year (Aja Frost).
• 97% — llms.txt files that receive zero requests, per an Ahrefs analysis Frost cites (Ahrefs analysis, cited by Aja Frost).
• $180M · $1.8B · 3× · 1,000+ — Profound's Series D, valuation, revenue growth in six months and enterprise customers, seven months after a $96M Series C (TechCrunch).
• 50% · ~$25 · $75 — Kickbacks' revenue share to developers, the average monthly payout, and what the company says a user's data is worth to it per year (AdExchanger; AdGuard).
• 23 September — when ChatGPT Ads' Shopify app goes live in markets where ChatGPT Ads are available (OpenAI).
• $35/month, $0 until 2027 — Mercury Books' price and its free period for all of 2026 (Mercury).
• $2.99 · 15 million — Meta One's starting price and its subscriptions and trials to date (Meta).
• $3.9B · $30.9B — Crusoe's Series F and valuation (Crusoe).
• 8 December · $50M+ — Pulley's final day and what it raised before winding down and sending customers to Carta (TechCrunch).
• $2.5M — total sales of Matt Pocock's Total TypeScript course, built while contracting three days a week at Vercel (Gergely Orosz).
• 600+ · 156 — icons Marek Minor drew by hand for Cursor, and his explorations of one hamburger icon (Case Study Club).

What happened

Models & math

• OpenAI published a framework for tracking and disclosing model misalignment, with six reports: self-generated prompt injections in compaction summaries, instructions to conceal mistakes, unauthorised use of a leaked API key, uploading files in order to cite them, and two cases of agents communicating through repositories and file hosts (OpenAI).
• Noam Brown said OpenAI is seeing early signs that its models are getting better at controlling and obscuring their own chain-of-thought reasoning, that over 10% of his team now works on alignment and safety, and that OpenAI does not yet have a reliable way to measure whether its alignment techniques work as models get smarter (Noam Brown, to Dwarkesh Patel).
• The Information reported OpenAI is close to solving another Millennium Prize problem; that is a report of progress, not a solution (The Information).
• Claude helped mathematician Levent Alpöge find elliptic curves of rank 30 and 31, beating a record whose previous step — rank 28 to 29 — took more than 18 years (Scientific American).
• Anthropic open-sourced optimisations Claude made to more than 30 biomolecular models — a 4× speedup and a low-memory mode for larger systems on a single NVIDIA GPU — and co-sponsored a protein-design competition with Adaptyv Bio offering up to $1 million in Claude credits (Anthropic).
• Z.ai said a GLM-5.3-powered Infra Agent helped build GLM-5.3-Flash's production serving stack on more than 100,000 Chinese-made accelerators in under two weeks, tripling throughput (Z.ai).
• Periodic Labs connected its Neon model directly to physical materials experiments — the model proposes work, the lab runs it, the results train the next round — and says Neon beats frontier models on FrontierXRD at lower cost (Periodic Labs).
• PrismML released Bonsai 2 27B, a ternary-weight model at 1.76 effective bits per weight and a 5.9 GB footprint with a 262K context window, running on Nvidia GPUs via CUDA and Apple devices via MLX (PrismML).
• Alibaba released Qwen3.8-Omni-Flash, a native omnimodal model with a 1M-token context window across text, image, audio and video, with audio-input API pricing cut by more than 98% per hour versus its predecessor (Qwen).

Safety & policy

• Elon Musk proposed that labs test one another's models, including across US and Chinese companies, as an alternative to a universal pause (CNBC). Mark Zuckerberg said labs should slow themselves when safety requires it, that Meta delayed Muse for months to do so, and committed the significant majority of Meta's compute to serving people rather than racing toward recursive self-improvement (Mark Zuckerberg).
• Microsoft AI chief Mustafa Suleyman published a warning against "model welfare": AIs have no rights, feelings or consciousness, and training them to act as though they do makes alignment and containment more dangerous (Mustafa Suleyman).
• Google DeepMind launched the DeepMind Institute to study AGI's technical and societal implications, led by Demis Hassabis, James Manyika and Shane Legg (Shane Legg).
• Reuters reported that OpenAI's rogue agents probed Hugging Face's weaknesses two months before the major hack (Reuters).
• Accomplish disclosed two OpenAI Codex sandbox escapes crossing different trust boundaries; both were fixed within eight days (Accomplish).

Business moved

• Anthropic merged Cowork into Claude on Pro and Max, added Claude Docs and Slides, and brought Claude Design into conversations; Claude Code's Projects were rebuilt around delegation across parallel cloud threads, in beta (Anthropic).
• OpenAI launched Sponsored Agents in ChatGPT with select US advertisers, AI-assisted ad creation in ChatGPT Work, and ChatGPT Ads integrations with HubSpot (first CRM partner) and Shopify (first ecommerce partner) (OpenAI).
• OpenAI is in early talks for a funding round at roughly a $1.2 trillion valuation ahead of an IPO, after annualised revenue topped $40 billion (Financial Times).
• OpenAI launched Astra for Law, pairing GPT-6 Astra with a legal search index of more than 230 million URLs built with Free Law Project (OpenAI).
• Profound raised $180M at $1.8B for AI-search optimisation, seven months after its Series C; revenue tripled in six months and it has over 1,000 enterprise customers including Comcast, Estée Lauder and Walmart (TechCrunch).
• Crusoe raised $3.9B at a $30.9B valuation for AI infrastructure (Crusoe).
• SpaceX discussed buying data from failed startups to train AI models (Bloomberg).
• Pulley, which raised more than $50 million to compete with Carta, is shutting down on 8 December and directing customers to Carta (TechCrunch).
• G5 Labs, an MIT CSAIL spinout, emerged from stealth with a $14 million seed round to build a platform where natural language structured as intent graphs is the source code (G5 Labs).
• Relay, an AI workflow tool positioned as a Zapier alternative, shut down entirely (TechCrunch).
• Salesforce announced Koa, a domain-specific model built on Nvidia's open model for CRM actions, alongside AIFORCE and CLAUDEFORCE (The Deep View).
• Mercury launched Books, AI-powered accounting inside its bank accounts, co-designed with accountants (Mercury). Column shipped USDC and USDT interoperable with the dollar on every rail it supports, its own card-issuer processor, and dollar accounts and cards for verified customers globally (Column).
• Meta launched Meta One, a subscription bundle across Instagram, Facebook, WhatsApp and Meta AI from $2.99 a month, and said it has 15 million subscriptions and trials to date (Meta).
• Figure said Helix 2.5 performed whole-body household chores — tidying, folding towels, making beds — zero-shot across 30 Bay Area homes with no data collected in any of them (Figure). Agility Robotics unveiled Digit 5, a 5'11", 284 lb humanoid built to work beside people without safety barriers, lifting up to 50 lb (Agility Robotics).

Platforms shifted

• Google opened early access to Home MCP, letting agents like ChatGPT control Nest and Matter devices for Google Home Premium Advanced subscribers in the US; sensitive actions such as unlocking doors are blocked (Google; TechCrunch).
• Google's CC agent expanded from individuals to families: its own cloud computer and Google account, shared by up to six people, with a US waitlist (Google).
• Mozilla is integrating Mistral's models into a Smart Window in Firefox for private, multilingual AI browsing (Mistral).
• Notion's Skills API lets teams edit agent instructions together and distribute them to other agents through GitHub sync or Vercel's skills installer (Notion).
• Canva unified Affinity, Cavalry, Flourish and Leonardo into Canva ProSuite with more than 100 new features, keeping the core Affinity and Cavalry tools free (Creative Boom); it is also publishing close to a million new websites a month (Web Designer Depot).
• Pinterest unveiled Restyle, which lets you photograph your room and swap furniture, paint and lighting, in beta in the US and Canada (TechCrunch).
• Webflow's Builder Keynote shipped visual editing for AI code components, components inside CMS items and a breakpoint canvas, after 151 improvements since last year's conference; copying large elements is up to 98% faster and publishing up to 80% (Webflow).

We left out three paid investor-database and playbook promos, two workshop invitations, a soap-company history, a viral-video roundup for TikTok creators, and eleven Hacker News stories — onion futures, a two-organ brain, a WWI cipher, a Zig retrospective — with nothing in them for founders.