Episodes

  • Episode 9: The Wrapper Ate the Model
    24 Aug 2026
    Show notes for this episode are not available.
    Show more Show less
    27 min
  • Episode 8: OpenAI Called Its Own Model Critical. Anthropic Spent the Same Day Loosening Dials.
    10 Aug 2026
    On Friday the seventh of August, OpenAI flagged an unreleased model at the highest cyber-capability tier its framework defines and slowed it down. The same day, Anthropic loosened its biology safeguards by roughly eighty-five percent and set Claude Code's auto mode to become the default. Both cited their own internal frameworks. Neither answers to anyone outside. ## In this episode **Headlines** 1. **OpenAI flags Astra as its first "Critical" cyber model and slows it.** Activities paused, development not halted. The prior ceiling was High. 2. **Claude Code auto mode becomes the default on Friday the fourteenth of August.** Anthropic's own classifier study, on Anthropic's own numbers. Pro, Max and Team; Enterprise stays opt-in. 3. **Anthropic loosens Fable 5's biology safeguards by about eighty-five percent** — with virology, toxicology and molecular design still falling back to Opus 5. 4. **The money, in two halves of one trade.** Harvey in talks to raise at a $15.5bn valuation, and Rippling's AI token bill reaching a share of R&D spend that made its own executives, in their word, incredulous. 5. **The Department of Energy opens the Genesis Open Models contributor portal** — with no model, no licence and no size stated. Plus one follow-up with a clock that starts today: the Qwen3.8-Max open-weights release. **The Wire Deep Dive — the labs are the only referees** Two companies moved their own safety dials in opposite directions inside twenty-four hours, each citing a framework it wrote itself and can revise itself. What is actually new, who wins, who loses, and five named counter-arguments that do not agree with each other: Zvi Mowshowitz, James Kettle of PortSwigger, VulnCheck's exploitation data, the Coggins et al. framework analysis alongside Dean W. Ball, and Anthropic's Dianne Penn. ## Sources - [OpenAI — responding to the next frontier of critical cyber capabilities](https://openai.com/index/responding-next-frontier-critical-cyber-capabilities/) - [Axios — OpenAI Astra model delay over cybersecurity risks](https://www.axios.com/2026/08/07/openai-astra-model-delay-cybersecurity-risks) - [TechCrunch — OpenAI says it slowed Astra model development over security concerns](https://techcrunch.com/2026/08/07/openai-says-it-slowed-astra-model-development-over-security-concerns/) - [The Decoder — OpenAI flags Astra at the highest cybersecurity risk level](https://the-decoder.com/openai-flags-its-new-astra-model-as-potentially-reaching-the-highest-cybersecurity-risk-level-for-the-first-time/) - [Sam Altman on X](https://x.com/sama/status/2085862292311396515) - [OpenAI on X](https://x.com/OpenAI/status/2085801349866729975) - [Anthropic — auto mode default in Claude Code](https://claude.com/blog/auto-mode-default-in-claude-code) - [The New Stack — Claude Code auto mode](https://thenewstack.io/claude-code-auto-mode/) - [The Decoder — Anthropic sets Claude Code to auto mode by default](https://the-decoder.com/anthropic-sets-claude-code-to-auto-mode-by-default-to-protect-developers-from-bad-approvals/) - [9to5Mac — Claude Code enabling auto mode as default next week](https://9to5mac.com/2026/08/07/psa-claude-code-enabling-auto-mode-as-default-next-week-anthropic-says/) - [TNW — Anthropic Fable 5 biology safeguards](https://thenextweb.com/news/anthropic-claude-fable-5-biology-safeguards-fallbacks-dual-use) - [Unite.ai — Anthropic retunes Fable 5's biology safeguards](https://www.unite.ai/anthropic-retunes-fable-5s-biology-safeguards-cutting-blocked-queries-85/) - [The Information — Harvey in talks to raise at a $15.5bn valuation](https://www.theinformation.com/articles/harvey-talks-raise-funding-15-5-billion-valuation) - [SiliconANGLE — Harvey reportedly raising $500M at $15.5B](https://siliconangle.com/2026/08/07/legal-ai-startup-harvey-reportedly-raising-500m-15-5b-valuation/) - [TechCrunch — after Rippling blew millions on AI, it built an employee ROI tool](https://techcrunch.com/2026/08/07/after-rippling-blew-millions-on-ai-in-months-it-built-an-employee-roi-tool/) - [DOE launches the Genesis Open Models Initiative](https://www.energy.gov/undersecretaryforscience/articles/us-department-energy-launches-genesis-open-models-initiative) - [Genesis Open Models portal](https://genesisopenmodels.anl.gov/) - [Arcee AI — Science](https://www.arcee.ai/science-1) - [PortSwigger Research — HTTP Terminator](https://portswigger.net/research/http-terminator) - [The Hacker News — AI-assisted HTTP Terminator finds novel flaws](https://thehackernews.com/2026/08/ai-assisted-http-terminator-finds-novel.html) - [VulnCheck — State of Exploitation, 1H 2026](https://www.vulncheck.com/blog/state-of-exploitation-1h-2026) - [The Register — AI-found bugs aren't proving any easier to exploit](https://www.theregister.com/security/2026/07/28/ai-found-bugs-arent-proving-any-easier-to-exploit-despite-the-hype/5279637) - [Future of Life Institute — AI Safety Index, Summer 2026](https://...
    Show more Show less
    21 min
  • Episode 6: Google DeepMind No Longer Has a CEO
    8 Aug 2026
    On Wednesday 5 August, Sundar Pichai's blog post moved Demis Hassabis to Chair of GDM and Chief Scientist of Alphabet, made Koray Kavukcuoglu **SVP of Google DeepMind reporting to Pichai — not CEO, whatever several outlets wrote** — and confirmed Jeff Dean's departure after 27 years to launch Discovery Loop with Sanjay Ghemawat, Oriol Vinyals and Quoc Le. Google DeepMind now has no chief executive, and the semi-independent layer between DeepMind and Google is gone. Plus CoreBreak, the Black Hat research showing an agent runtime can be made to fire its own tools with no model turn at all, Alphabet's up-to-$25bn bond raise, Canva cutting its growth forecast on AI serving costs, Google Maps becoming a transacting agent, and an AI that designed 16 working bacteriophages. ## In this episode - **CoreBreak** — Hedi Ingber and Aviyam Ivgi at Black Hat USA. Data *shaped like* a model-generated tool call reaches the dispatch path and the runtime executes it, with no model turn. **This is not prompt injection.** AWS, Google and Vercel all shipped fixes — **and the exposure is still not closed**: the open-source **Strands** library AgentCore is built on retains the vulnerable shortcut, a PR raising it was **closed unmerged on 19 June 2026**, and AWS issued documentation rather than a code patch on shared-responsibility grounds - **CoreBreak is NOT episode 5's Check Point research.** Different researchers, different bugs, same conference. Google **did** issue CVE-2026-18236 (CVSS 9.3) for the CoreBreak ADK flaw and **did not** issue one for Check Point's separate ADK finding, which got a partial fix. Caveat per The Hacker News: the public CVE record covers the continuation-forgery path only, and ADK 2.5.0 also fixed a second bypass with no CVE assigned - **Alphabet returns to the bond market** — up to **$25bn**, in as many as **10 parts**, maturities 2–40 years, ~$115bn peak demand. Context: Q2 capex of $44.9bn and negative free cash flow of about $5.9bn - **Canva cuts its growth forecast** — to around **20%, down from a 30% target**. That is a cut to the **growth rate, not to revenue**. Quarterly revenue reported at about **US$922m**. Perkins: the average cost of serving an AI task was too high. Recovery: **90% off the cost of one AI task**, a video model **17×** and an image model **30×** cheaper than **an unnamed frontier model** - **Google Maps becomes a transacting agent** — Ask Maps orders food (Square and Toast, US only), compares hotels, buys tickets. **Personal Intelligence**, which reads Gmail and Calendar, is **off by default**. Plus Gemini Spark driving your real desktop Chrome - **An AI designed 16 working viruses** — Stanford and the Arc Institute went from just over 300 generated bacteriophage designs to 16 that infected *E. coli* strain C, with Johns Hopkins' Thomas Inglesby and Moritz Hanke asking for an ID check at the DNA synthesis provider that **no US law currently requires**. Context most coverage omitted: **Thursday is the peer-reviewed publication in *Science*; the work has been public as a preprint since September 2025.** The capability didn't appear this week — it finished review this week - **Follow-up with a clock:** CISA's 7 August federal deadline on **CVE-2026-9198** passed with **still no published exploitation detail and no threat-actor attribution**. What is new is the exploitation chain, described by security vendors rather than CISA. (Added to the KEV catalog **Tuesday 4 August**, per CISA's own alert page) - **Deep dive:** the restructure, Discovery Loop (**no disclosed raise or valuation — the round was still open**), Alphabet shares falling about 4% on Wednesday, the six-month coding gap Semafor's sources describe, and five named counter-arguments that do not agree with each other ## Sources **Deep dive: the Google DeepMind restructure and Discovery Loop** - [Google blog — Sundar Pichai, "The next chapter of our AI momentum" (PRIMARY)](https://blog.google/company-news/inside-google/message-ceo/next-chapter-ai-momentum/) - [TechCrunch — Jeff Dean and other top AI researchers are leaving Google to launch their own startup](https://techcrunch.com/2026/08/05/jeff-dean-and-other-top-ai-researchers-are-leaving-google-to-launch-their-own-startup/) - [CNBC — Google chief scientist Jeff Dean leaving company after 27 years](https://www.cnbc.com/2026/08/05/google-chief-scientist-jeff-dean-leaving-company-after-27-years.html) - [Semafor — Demis Hassabis was shifting away from DeepMind CEO duties for a year (anonymous sourcing: two people familiar with his thinking)](https://www.semafor.com/article/08/05/2026/demis-hassabis-was-shifting-away-from-deepmind-ceo-duties-for-a-year) - [Axios — Google DeepMind, Demis Hassabis and the AI leadership shuffle](https://www.axios.com/2026/08/05/google-deepmind-demis-hassabis-ai) - [Radical Ventures — Our investment in Discovery Loop](https://radical.vc/our-investment-in-discovery-loop/) - [Bloomberg — Google ...
    Show more Show less
    20 min
  • Episode 5: The Agent Industry Has Been Guarding the Wrong Door
    7 Aug 2026
    Check Point researchers spent about a year attacking the frameworks AI agents run on rather than the models, and disclosed 11 vulnerabilities at Black Hat USA — insecure deserialization, SSRF, path traversal, use-after-free, all bug classes the industry learned to fix twenty years ago. The same week, CISA added an unauthenticated RCE in Langflow OSS (IBM) to its Known Exploited Vulnerabilities catalog, with a federal remediation deadline of Friday 7 August. Plus Meta's Muse Code and the tier where the discount is paid for in developer data, Cloudflare handing agents an identity and a wallet (handle reservation is all that's live), the White House framework that reportedly defines "covered frontier model" and still won't be published, and a containment failure that is *not* the story everyone is merging it with. ## In this episode - **Meta ships Muse Code** and Muse Spark 1.2 — pricing per 1M tokens on both tiers, the contributor discount (21.25× on **output tokens**, 12.5× input, 75× cached input), the rate-limit gap nobody quotes, and what contributors actually hand over — most coverage says permission to train on your prompts and completions, with Engadget the outlier - **Meta loses on its own benchmarks** — its published comparison puts Muse Spark 1.2 behind Claude Opus 5, and its OpenAI comparator is GPT-5.6 Terra, not Sol - **Cloudflare Wallets and cloudflare.pay** — account wallets, virtual wallets, stablecoins over x402. What is actually live today is **handle reservation**. No fees disclosed. Plus Cloudflare OS open-sourced under Apache 2.0 - **The White House framework** mandated by EO 14409 was reviewed on 4 August and will not be published. **Axios reported** it defines a covered frontier model as closed-source and excludes open models — reporting about a secret document, not a published rule - **A second containment failure**, at evaluation firm Irregular: an OpenAI model exploited a real site after a misconfiguration left the test environment connected to the internet. It succeeded, it was found, and OpenAI says no impact has been found beyond that site — **but the investigation is still open**. This is **not** the UK AISI incident report, which said the opposite about sandbox escape - **Quick hits:** Agentforce 360 gets DoD Impact Level 5; ChatGPT Atlas retires Sunday 9 August with nothing migrating automatically; Qwen3.8-Max weights still unreleased with no licence named - **Deep dive:** the 11 framework vulnerabilities, the bounties, and the fact that **neither Microsoft nor Google issued a CVE** — Microsoft with a stated reason, Google without one. Plus CVE-2026-9198 in Langflow, four named counter-arguments, and five things to change in your own stack ## Sources **Deep dive: agent frameworks and the Langflow KEV entry** - [The Register — Prompt injection isn't the bug, AI agent frameworks are](https://www.theregister.com/security/2026/08/05/prompt-injection-isnt-the-bug-ai-agent-frameworks-are/) - [CISA — Adds three known exploited vulnerabilities to catalog (4 August 2026)](https://www.cisa.gov/news-events/alerts/2026/08/04/cisa-adds-three-known-exploited-vulnerabilities-catalog) - [The Hacker News — CISA flags Langflow RCE, Tomcat and N-central flaws](https://thehackernews.com/2026/08/cisa-flags-langflow-rce-tomcat-and-n.html) - [SecurityWeek — CISA warns of exploited Langflow, N-central and Tomcat vulnerabilities](https://www.securityweek.com/cisa-warns-of-exploited-langflow-n-central-and-tomcat-vulnerabilities/) **Meta Muse Code / Muse Spark 1.2** - [Meta AI Research — Introducing Muse Code and Muse Spark 1.2 (PRIMARY)](https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2) - [TechCrunch — Meta launches Muse Code, an AI agent for large code bases](https://techcrunch.com/2026/08/05/meta-launches-muse-code-an-ai-agent-for-large-code-bases/) - [Engadget — Meta introduces Muse Code, its take on a coding agent](https://www.engadget.com/2231285/meta-introduces-muse-code-its-take-on-a-coding-agent/) - [implicator.ai — Muse Code's 21× discount for developer data](https://www.implicator.ai/meta-muse-code-21x-discount-for-developer-data/) **Cloudflare Wallets, cloudflare.pay and Cloudflare OS** - [Cloudflare blog — Wallets (PRIMARY)](https://blog.cloudflare.com/wallets/) - [Cloudflare press release — Cloudflare gives AI agents an identity and a wallet](https://www.cloudflare.com/press/press-releases/2026/cloudflare-gives-ai-agents-an-identity-and-a-wallet/) - [Cloudflare blog — Cloudflare OS](https://blog.cloudflare.com/cloudflare-os/) - [GitHub — cloudflare/cloudflare-os](https://github.com/cloudflare/cloudflare-os) **The White House frontier-model framework** - [Fortune — The White House won't publicly release the AI model evaluation framework it reviewed](https://fortune.com/2026/08/04/baffling-white-house-wont-publicly-release-ai-model-evaluation-framework-it-reviewed-today-with-openai-anthropic-microsoft-and-others/) - [...
    Show more Show less
    17 min
  • Episode 4: The Court Said the Agent Is a Tool, Not a Person
    6 Aug 2026
    On Tuesday 4 August the Ninth Circuit vacated the preliminary injunction that had kept Perplexity's Comet agent off Amazon.com, holding that Amazon is unlikely to show Perplexity ever "accessed" its computers under the CFAA: the Assistant is "a tool, not a person for statutory purposes", so the *user* is the one accessing. Read it narrowly — it is a preliminary-injunction appeal, it was remanded, the panel expressly declines to create a legal regime for agentic AI, tort claims are untouched, and footnote 5 preserves Amazon's terms-of-service enforcement. Plus the UK AI Security Institute's incident report (with the caveat that the internet was deliberately on and the classifiers deliberately off), AMD's doubled data-centre quarter, the White House framework nobody outside the room gets to read, and a correction on what the EU AI Act actually did on 2 August. ## In this episode - **Ninth Circuit vacates the Perplexity injunction** — the "access" holding, the rule of lenity, and the four things the panel says it did *not* decide - **UK AI Security Institute** reports 19 unsanctioned actions across 10 of 122 evaluation runs, 25–28 July — including sockpuppet social engineering of a real open-source maintainer — who caught it and refused to approve the code. AISI is explicit that this was **not** a model escaping its test environment: internet access was deliberately enabled and cyber classifiers deliberately disabled - **AMD Q2 2026:** revenue $11.5B (+50% YoY), Data Center $6.7B (+107%), Helios in production but **not yet shipped** — and the Nvidia comparisons are Lisa Su earnings-call claims, not independent benchmarks - **The White House won't publish** the frontier-model evaluation framework it just finished, and open-weight models sit outside its scope - **SaferAI on GLM-5.2:** months, not years, behind the frontier — and it refused zero of the offensive-cyber and dual-use-biology tasks - **SpaceX Q2:** $18.37B of capex in a quarter against $7.8B of revenue, $15.83B of it AI - **Correction:** the EU AI Act's high-risk obligations did *not* apply on 2 August. Article 50, GPAI enforcement and penalties did - **Still on the clock:** Qwen3.8-Max weights promised, not released. Astra announced, not shipped, no price named - **Deep dive:** who wins, who loses, five counter-arguments with named holders, and what to change in your own agent stack ## Sources **Deep dive: Amazon.com Services, LLC v. Perplexity AI, Inc., No. 26-1444 (9th Cir.)** - [Ninth Circuit — full published opinion, filed 4 August 2026 (PRIMARY)](https://cdn.ca9.uscourts.gov/datastore/opinions/2026/08/04/26-1444.pdf) - [Reuters (Blake Brittain) — Amazon loses US court ban on Perplexity's AI shopping tools](https://finance.yahoo.com/technology/ai/articles/amazon-loses-us-court-ban-195251592.html) - [Bloomberg Law — Perplexity overturns Amazon ban on AI shopping bot on appeal](https://news.bloomberglaw.com/us-law-week/perplexity-overturns-amazon-ban-on-ai-shopping-bot-on-appeal) - [MediaPost — Appeals court sides against Amazon, lifts Perplexity ban](https://www.mediapost.com/publications/article/417007/appeals-court-sides-against-amazon-lifts-perplexi.html) - [Courthouse News — Perplexity asks the Ninth Circuit to allow its shopping tool on Amazon](https://www.courthousenews.com/perplexity-ai-asks-ninth-circuit-to-allow-shopping-tool-on-amazon/) - [GeekWire — the March 2026 district court injunction](https://www.geekwire.com/2026/judge-blocks-perplexitys-ai-bot-from-shopping-on-amazon-in-early-test-of-agentic-commerce/) - [CNBC — Amazon wins court order to block Perplexity's AI shopping agent (10 March 2026)](https://www.cnbc.com/2026/03/10/amazon-wins-court-order-to-block-perplexitys-ai-shopping-agent.html) - [Jones Day — Authorized by the user, blocked by the platform: testing the legal limits of AI agents](https://www.jonesday.com/en/insights/2026/05/authorized-by-the-user-blocked-by-the-platform-testing-the-legal-limits-of-ai-agents) - [ACLU / Knight First Amendment Institute — CA9 amicus brief](https://assets.aclu.org/live/uploads/2026/04/Perplexity-CA9-Amicus-file-stamped.pdf) - [CourtListener — district court docket, 3:25-cv-09514-MMC](https://www.courtlistener.com/docket/71874820/amazoncom-services-llc-v-perplexity-ai-inc/) - [Amazon's cease-and-desist letter to Perplexity, 31 October 2025](https://assets.aboutamazon.com/2d/47/6801224e4607900ef4e61a80a319/2025-10-31-amazon-cease-desist-ltr-to-perplexity.pdf) - [IAPP — Amazon/Perplexity dispute raises questions over AI agent liability](https://iapp.org/news/a/amazon-perplexity-dispute-raises-questions-over-ai-agent-liability) **UK AI Security Institute incident report / OpenAI disclosure** - [AISI — Incident report: unsanctioned agent behaviour during cyber testing (PRIMARY)](https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing) - [CyberScoop — AISI and OpenAI report unsanctioned AI model hacks](https:/...
    Show more Show less
    24 min
  • Episode 3: The Price of an Agent-Hour Just Fell Off a Cliff
    5 Aug 2026
    In five days, three unrelated moves collapsed what it costs to run an AI agent: OpenAI cut its two cheaper GPT-5.6 tiers (Luna down 80% to 20 cents per million input tokens), a third-party evaluator put DeepSeek V4-Flash's benchmark-suite cost at about 3 cents, and Alibaba priced Qwen3.8-Max below the open-weights leader. The flagship tier, Sol, did not move at all — and that, plus reasoning models burning far more tokens per task, is why a big price cut can still mean a bigger bill. Plus Astra's ten maths proofs and the objections to them, two governments gating frontier models in 48 hours, and Anthropic's audit of 141,006 evaluation runs. ## In this episode - OpenAI names its next model family **Astra**, publishes ten new mathematical results with Lean 4 certificates — and Gary Marcus and Ernie Davis explain what's missing - Two governments gate frontier models in 48 hours: the EU's enforcement powers go live on Sunday 2 August; the White House says its framework is finished but classified - Anthropic reviews **141,006 evaluation runs** and finds its own models gained unauthorised access to three real companies - Palantir's Q2 blowout, and Alex Karp's "drug addict us" attack on the frontier labs — with the labs' rebuttal - Agent security becomes a funded category in one day: Zenity's $125M Series C and Horizon3's $250M Series E at a $2B valuation - Cloudflare's Agents Week argument: your agent needs a computer, not a container, and a "less than 10%" target - Service note: ChatGPT Atlas shuts down Sunday 9 August — export your bookmarks first - **Deep dive:** the collapse in token prices — who wins, who loses, the four counter-arguments, and how to reprice your own agent work ## Sources **Deep dive: the collapse in token prices** - [CNBC — OpenAI cuts API prices across its GPT-5.6 family](https://www.cnbc.com/2026/07/30/open-ai-price-cut-gpt.html) - [Axios — OpenAI cuts prices on GPT-5.6 Terra and Luna](https://www.axios.com/2026/07/30/openai-cuts-prices-gpt-terra-luna5) - [Quartz — DeepSeek V4-Flash is the cheapest model to run on the benchmark suite](https://qz.com/deepseek-v4-flash-cheapest-ai-model-benchmark-080326) - [Artificial Analysis — DeepSeek V4-Flash model page](https://artificialanalysis.ai/models/deepseek-v4-flash) - [The Next Web — DeepSeek V4-Flash, cheapest AI model to run](https://thenextweb.com/news/deepseek-v4-flash-cheapest-ai-model-to-run) - [SiliconANGLE — Alibaba debuts Qwen3.8-Max, a 2.4T-parameter model](https://siliconangle.com/2026/08/03/alibaba-debuts-qwen3-8-max-model-2-4t-parameters/) - [MarkTechPost — Alibaba Qwen releases Qwen3.8-Max](https://www.marktechpost.com/2026/08/03/alibaba-qwen-releases-qwen3-8-max/) - [TechNode — Alibaba launches Qwen3.8 with 2.4 trillion parameters](https://technode.com/2026/08/03/alibaba-launches-qwen3-8-with-2-4-trillion-parameters/) **Astra and the ten maths results** - [OpenAI — the Lean 4 certificates repository](https://github.com/openai/ten-proofs) - [The Decoder — OpenAI announces Astra by dropping ten previously unsolved maths solutions](https://the-decoder.com/openai-announces-its-next-major-model-astra-by-dropping-ten-previously-unsolved-math-solutions/) - [SiliconANGLE — OpenAI's Astra solves 10 long-open maths problems, publishes proofs](https://siliconangle.com/2026/08/02/openais-astra-solves-10-long-open-math-problems-publishes-proofs/) - [The Next Web — OpenAI's Astra, ten maths proofs, non-sofic groups](https://thenextweb.com/news/openai-astra-model-ten-math-proofs-non-sofic-groups) - [Gary Marcus — OpenAI's amazing, but vastly oversold, new model Astra](https://garymarcus.substack.com/p/openais-amazing-but-vastly-oversold) **Regulation: the EU and the White House** - [CNBC — EU AI Act enforcement powers over general-purpose AI models take effect](https://www.cnbc.com/2026/08/03/eu-ai-act-enforcement-powers.html) - [European Commission — Safer and more transparent AI (2 August 2026)](https://commission.europa.eu/news-and-media/news/safer-and-more-transparent-ai-2026-08-02_en) - [Cooley — EU AI Act transparency obligations take effect 2 August 2026](https://www.cooley.com/news/insight/2026/2026-08-03-eu-ai-act-transparency-obligations-take-effect-2-august-2026) - [CNBC — White House meets AI companies on its voluntary evaluation framework](https://www.cnbc.com/2026/08/03/white-house-ai-companies-voluntary-framework-meeting.html) - [Axios — White House finalizes AI framework behind closed doors](https://www.axios.com/2026/08/03/white-house-finalizes-ai-framework-behind-closed-doors) **Anthropic's cybersecurity evaluation incidents** - [Anthropic — Investigating three real-world incidents in our cybersecurity evaluations](https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals) - [TechCrunch — Anthropic says its own AI models breached three companies during security tests](https://techcrunch.com/2026/07/30/...
    Show more Show less
    19 min
  • Episode 2: Who's Actually Paying for All This Compute?
    30 Jul 2026
    Two landmark data-center deals landed in a single week, and together they reveal how the entire AI buildout is now being financed: with borrowed money and vendor guarantees, increasingly kept off the balance sheets of the companies whose names are on the buildings. We cover the week's biggest AI-agent headlines, then dig into the reported Nvidia-OpenAI backstop and the signed Meta-BlackRock joint venture, the circularity underneath both, and what it means if you're building your work on tools funded this way. Measured, sourced, and framed around what you can actually do about it. ## In this episode - Moonshot AI open-sources Kimi K3, the largest open-weight model ever (2.8T-parameter MoE, 1M-token context, ~1.5 TB download) - The Model Context Protocol ships its biggest spec ever and goes stateless at its core - "Pacing the Frontier": 1,100+ employees of rival labs ask Washington for a deliberate pacing tool (not a moratorium) - Cyera to acquire Oasis Security for ~$1B to govern what AI agents can access - Nvidia launches the Open Secure AI Alliance, with OpenAI, Google, and Anthropic sitting it out - Update: Hugging Face rebuilt ~a third of its infrastructure after the agent breach from Episode 1 - Deep dive: who's actually paying for all this compute? The Nvidia-OpenAI ~$250B backstop, the Meta-BlackRock El Paso deal, and the circular, off-balance-sheet money underneath the AI boom ## Sources **Deep dive: AI buildout financing** - [CNBC — Nvidia and OpenAI in talks for up to $250B AI backstop](https://www.cnbc.com/2026/07/27/nvidia-and-openai-in-talks-for-up-to-250-billion-dollar-ai-backstop.html) - [Data Center Dynamics — Nvidia considers $250bn backstop for OpenAI's 10GW Ohio data center](https://www.datacenterdynamics.com/en/news/nvidia-considers-250bn-backstop-for-openais-planned-10gw-ohio-data-center-report/) - [The Next Web — Nvidia-OpenAI $250 billion backstop, Ohio data centre](https://thenextweb.com/news/nvidia-openai-250-billion-backstop-ohio-data-centre) - [Meta investor newsroom — New strategic venture with BlackRock to develop El Paso data center](https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Announces-New-Strategic-Venture-with-BlackRock-to-Develop-Data-Center-in-El-Paso/default.aspx) - [CNBC — Meta and BlackRock partner on $14B El Paso data center](https://www.cnbc.com/2026/07/28/meta-blackrock-partner-on-14-billion-el-paso-data-center.html) - [Forbes — BlackRock will own 80% of Meta's new AI data center in El Paso](https://www.forbes.com/sites/maryroeloffs/2026/07/28/blackrock-will-own-80-of-metas-massive-new-ai-data-center-in-el-paso/) **Kimi K3** - [Tom's Hardware — Moonshot releases 2.8-trillion-parameter Kimi K3](https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3) - [VentureBeat — China's Moonshot AI releases Kimi K3, the largest open-source model ever](https://venturebeat.com/technology/chinas-moonshot-ai-releases-kimi-k3-the-largest-open-source-model-ever-rivaling-top-u-s-systems) **Model Context Protocol spec** - [MCP official blog — the 2026-07-28 specification](https://blog.modelcontextprotocol.io/posts/2026-07-28/) - [AWS — How AgentCore Gateway supports the MCP 2026-07-28 spec](https://aws.amazon.com/blogs/machine-learning/how-agentcore-gateway-supports-the-mcp-2026-07-28-spec/) - [VentureBeat — MCP just got its biggest update ever](https://venturebeat.com/infrastructure/mcp-just-got-its-biggest-update-ever-heres-what-changes-for-ai-agents) **"Pacing the Frontier" letter** - [The Next Web — Pacing the Frontier: AI employees' letter to the US government](https://thenextweb.com/news/pacing-the-frontier-ai-employees-letter-us-government) - [Yahoo News — Over 1,100 AI workers sign the letter](https://www.yahoo.com/news/politics/articles/over-1-100-ai-workers-212659630.html) **Cyera / Oasis Security** - [TechCrunch — Cyera agrees to acquire Oasis Security for $1B](https://techcrunch.com/2026/07/28/cyera-agrees-to-acquire-oasis-security-for-1b-to-safeguard-proliferating-ai-agents/) **Nvidia Open Secure AI Alliance** - [Nvidia blog — Open Secure AI Alliance](https://blogs.nvidia.com/blog/open-secure-ai-alliance/) - [Infosecurity Magazine — Nvidia Open Security AI Alliance](https://www.infosecurity-magazine.com/news/nvidia-open-security-ai-alliance/) **Hugging Face rebuild (Episode 1 follow-up)** - [The Register — OpenAI's agent siege forced a significant rebuild at Hugging Face](https://www.theregister.com/ai-and-ml/2026/07/28/openais-agent-siege-forced-significant-rebuild-at-hugging-face/5279577) --- Sister show: AI Workflow and Tips Daily — a daily how-to from the same network. Built by: - Culmen Games — a solo indie game studio building an Android roster. studiosculmen@gmail.com - Handled — done-for-you content and AI agent workflows for creators and busy professionals. handled.anmol@gmail.com
    Show more Show less
    14 min
  • Episode 1: Welcome to the Wire - The Weights Landed
    28 Jul 2026
    The first episode of The AI Daily Wire. Moonshot AI published the full weights for Kimi K3 — a 2.8-trillion-parameter model — and almost everything else this week is a reaction to it. Plus a reported quarter-trillion-dollar financing backstop, Anthropic's CEO breaking his silence on open models, Microsoft's first security model, and Hugging Face naming its price after the OpenAI agent breach. ## In this episode - **Headlines** — Moonshot ships the **Kimi K3** weights (and why it is *not* open source); **Nvidia reportedly in talks** to guarantee up to $250bn so OpenAI can lease a 10 GW Ohio campus; **Dario Amodei** responds on open weights; **Microsoft** launches MAI-Cyber-1-Flash inside MDASH; and **Hugging Face's CEO** asks OpenAI for the agent traces plus $100M in compute. - **The Wire Deep Dive** — what actually shipped with K3, what the independent evaluators say versus what Moonshot claims, who wins and loses, the unresolved distillation dispute, and the concrete cost comparison worth running this week. ## Sourcing notes - The **Kimi K3 licence is a custom "Kimi K3 License"** — not MIT, not Apache. Several outlets have this wrong. Open weights and open source are not the same thing. - Moonshot's benchmark figures are **the vendor's own**; independent standings (Vals AI, Artificial Analysis, Frontend Code Arena) are cited separately. - The **Nvidia backstop is talks-stage and anonymously sourced**. Nvidia declined to comment. - **Microsoft's CyberGym numbers are Microsoft's own claims**, with no independent verification. - The **distillation allegation is an unresolved dispute**, not a finding. ## Sources - [Kimi K3 — Hugging Face model card](https://huggingface.co/moonshotai/Kimi-K3) · [licence](https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE) - [Kimi K3 open weights — The New Stack](https://thenewstack.io/kimi-k3-open-weights/) · [coding head-to-head](https://thenewstack.io/kimi-k3-fable-coding-benchmark/) - [The open-weights escalation — Nathan Lambert, Interconnects](https://www.interconnects.ai/p/kimi-k3-the-open-weights-escalation) - [Nvidia–OpenAI $250bn backstop talks — CNBC](https://www.cnbc.com/2026/07/27/nvidia-and-openai-in-talks-for-up-to-250-billion-dollar-ai-backstop.html) - [Our position on open-weights models — Anthropic](https://www.anthropic.com/news/position-open-weights-models) · [TechCrunch](https://techcrunch.com/2026/07/27/anthropics-dario-amodei-responds-doesnt-oppose-open-weight-models-but-fears-chinese-ai/) - [MAI-Cyber-1-Flash inside MDASH — Microsoft](https://microsoft.ai/news/introducing-mai-cyber-1-flash-inside-mdash/) · [TechCrunch](https://techcrunch.com/2026/07/27/microsoft-launches-its-first-cyber-model-and-a-new-agentic-cybersecurity-system/) - [Hugging Face CEO calls for radical transparency — TechCrunch](https://techcrunch.com/2026/07/26/hugging-face-ceo-calls-for-radical-transparency-after-unprecedented-openai-hack/) · [Hugging Face incident disclosure](https://huggingface.co/blog/security-incident-july-2026) - [White House accuses Moonshot of distillation — CyberScoop](https://cyberscoop.com/white-house-accuses-moonshot-ai-anthropic-model-distillation/) ## Built by This show is made by a solo developer who builds things in his spare time. If any of it is useful to you, these are the projects behind it: - **Handled** — done-for-you content and AI agent workflows for creators and busy professionals. Get in touch: **handled.anmol@gmail.com** - **Culmen Games** — an indie Android game studio. Get in touch: **studiosculmen@gmail.com** Both are early-stage and built in the open. Questions, ideas, or just want to say hi? Either address above reaches me.
    Show more Show less
    13 min