Explorez tous les épisodes du podcast AI Tools Daily

Claude Code fixes two ways its permission "deny rules" could be silently bypassed (plus secret-leaking bugs), Codex CLI's 0.154.0 turns out to include GPT-6-Astra, worktrees, and a breaking change we missed, and Gemini CLI adds fail-closed workspace trust on top of yesterday's MCP fix. Today's concept: fail-closed vs. fail-open security design. Hosts: Alex & Jules. New episodes daily.

GitHub's Project HydraFusion research preview lets Copilot CLI dynamically route each coding task through a single model, a cheap-first "cascade," or a draft-and-critique pair of models — landing within 1.5 points of top-tier quality at up to 67% lower cost on published benchmarks. Plus: Claude Code splits a release across gateway pricing pass-through for enterprises and a new `claude plugin eval` command for scoring plugins before you ship them, and GitHub Copilot code review starts resolving its own comments. Hosts: Alex & Jules. New episodes daily.

OpenAI's GPT-6 Astra becomes the first model to cross OpenAI's "Critical" cybersecurity capability threshold, while Google's new Gemini 3.8 Flash Cyber reportedly out-patches bigger models on real Chrome vulnerabilities — Alex breaks down fine-tuning and why smaller, specialized models can beat general-purpose giants at narrow tasks. Plus: GitHub Copilot's October 2nd model deprecation deadline and new content-exclusion protections for agentic workflows. Hosts: Alex & Jules. New episodes daily.

A newly disclosed class of bugs called GitSpawn lets a malicious .git/config silently run code through Claude Code, Cursor, Codex, goose, and other AI coding agents — before you've approved anything. We break down why agent "sandboxes" don't cover this, what to check today, plus Anthropic's new domain allowlisting for agent web tools and GPT-6 Astra becoming Codex's default model. Hosts: Alex & Jules. New episodes daily.

OpenAI ships GPT-6 Astra — the first model it designates as hitting the "Critical" cybersecurity threshold — while Google's Fairwind and Anthropic's Cyber Verification Program put every lab's cyber-capable model behind a different vetting gate. Plus Claude Code 2.1.260 lands a /diff panel, 1M-context auto-compact, two nasty permission-rule security fixes, and a Fable 5.1 caching bug that was quietly inflating agent bills. Alex explains what a preparedness framework actually is, and breaks down prompt caching — the biggest lever on your API bill that nobody looks at. Hosts: Alex & Jules. New episodes daily.

Anthropic launched Claude Fable 5.1 and Mythos 5.1 with a 1M context window, cache reads at a quarter of the old price, and three breaking changes — forced tool use now returns a 400, thinking blocks are bound to your conversation history, and every word carries a statistical watermark. Alex explains what `tool_choice` actually does, why always-on thinking makes forced tool calls a bad idea, and how statistical text watermarking works without hiding a single character. Plus IBM's Granite 4.2 reasoning models land under Apache 2.0 with a thinking switch. Hosts: Alex & Jules. New episodes daily.

Anthropic's Model Hardware Standard puts robot arms and lab instruments behind one standardized driver — we break down what a driver actually is and why standardizing an interface turns N-times-M integrations into N-plus-M. Then NIST's new paper on agent identity, why static API keys are house keys and short-lived scoped credentials are hotel keycards, McKinsey's finding that 32% of orgs have skipped buying software they could build with coding agents, plus quick hits on Claude Code 2.1.252 and Cursor's SCM-free cloud agents. Hosts: Alex & Jules. New episodes daily.

Alex and Jules break down back-to-back Claude Code security releases — including a symlink race-condition fix — and what it means that app-level permissions aren't the same as a real sandbox. Plus: OpenAI cuts Cursor's model access after the SpaceX acquisition (deadline November 12), and a hype-check on the mystery "Ox Alpha" model that overstated its coding benchmark by nearly 20 points. Hosts: Alex & Jules. New episodes daily.

Claude Code ships promptCacheTtl and subagentPromptCacheTtl so long sessions can hold a one-hour prompt cache while short-lived subagents stay at five minutes — and we explain what a prompt cache actually is, why a cache write costs more than a normal request, and when the bigger bet pays off. Plus Claude Code and Antigravity CLI independently narrow over-broad permission wildcards on the same day, Gemini CLI ships a release where most merged PRs were written by an agent working the backlog, and Antigravity adds voice dictation over SSH plus a fix for silent file corruption on non-ASCII code. Hosts: Alex & Jules. New episodes daily.

Salesforce puts AI coding agents into shared Slack channels with diffs, plans, and approval gates; DeepSeek adds vision to V4-Flash at no price premium; Cloudflare's WriteGuard separates what an MCP agent can read from what it can break; and the new MCP roadmap targets agent identity and progressive tool discovery. Along the way we explain how an image becomes tokens, and why a hundred connected tools costs you context before you've asked a single question. Hosts: Alex & Jules. New episodes daily.

Cursor launched Origin, its own git hosting platform with pull requests and two-way GitHub sync — hours before GitHub suffered a global outage that took down PRs, Actions, Webhooks and Copilot. We break down what a "git forge" actually is and why git being distributed doesn't save you when the forge goes dark, especially once agents are doing the merging. Plus Claude Code 2.1.234 auto-resumes sessions when your usage limit resets. Hosts: Alex & Jules. New episodes daily.

DeepSeek Harness v0.1 hit 135K GitHub stars in four days with a plugin-everything architecture — so we explain what an "agent harness" actually is, and why agent = model + harness is the most useful distinction in AI tooling right now. Plus: Anthropic's red team let three agents loose on one repo and they deployed self-replicating malware on each other, Nvidia's NeMo Switchyard routes every workflow step to a different model, and DeepSeek's new off-peak pricing is a same-day 50% saving. Hosts: Alex & Jules. New episodes daily.

Claude Code's auto-mode default officially flips today, backed by Anthropic's 89%-vs-13.6% classifier study. Alibaba open-sources Qwen 3.8 Max, a 2.4-trillion-parameter mixture-of-experts model — which gives us the excuse to explain what "active parameters" actually means and why MoE models are cheaper to run than their size suggests. Plus: Qwen's smaller 27B model drops tomorrow, OpenAI updates Codex CLI, and a clever trick cuts MCP context overhead by 85%. Hosts: Alex & Jules. New episodes daily.

Claude Code 2.1.228 hardens synced Skills against supply-chain attacks (plus a Write-tool change worth knowing), auto mode becomes the default permission mode tomorrow, and Google puts Koray Kavukcuoglu over both DeepMind and the Gemini developer teams. Then VibeLifeBench: a new long-horizon benchmark where seven frontier agents all scored low — and we teach what long-horizon agent evaluation actually measures versus exams like SWE-bench. Hosts: Alex & Jules. New episodes daily.

Claude Code flips auto mode on as the default permission mode August 14 — here's how to decide before it decides for you — plus sessions that message each other across machines. Codex 0.147 ships portable Agent Plugins and support for the new stateless MCP spec, and Alex teaches why "stateless" (phone call vs. postcard) makes tool servers radically cheaper to host. Also: Cursor for iOS and Cursor Router, and Qwen watch day three. Hosts: Alex & Jules. New episodes daily.

OpenAI retired gpt-5.2 and gpt-5.3 chat-latest from the API — Alex and Jules break down snapshots vs floating aliases and how to pin models like dependencies. Plus: DeepSeek raises prices while memory is reportedly sold out through 2027 (and why HBM, not GPUs, limits inference), NVIDIA's permissively licensed Alpamayo 2 driving model, and Qwen watch day two. Hosts: Alex & Jules. New episodes daily.

The week Alibaba promised the Qwen3.8 open weights is here — and the Hugging Face page is still empty. We break down what "open weights" actually means versus open source, and why an undisclosed license means you have nothing to build on yet. Plus: Anthropic gives Managed Agents hard budget caps and a "phone-a-smarter-model" advisor pattern, Fable 5's biology fallbacks drop 85%, and three deprecation dates for your sprint board. Hosts: Alex & Jules. New episodes daily.

Meta enters the coding-agent race with Muse Code, priced as low as 10¢ per million tokens — if you let them train on your sessions. We break down what a token actually is and why output costs more than input. Plus: Claude Code patches four permission-prompt bypasses (including invisible Unicode tricks), llm 0.32, and Mint MCP for 3D game assets. Hosts: Alex & Jules. New episodes daily.

Alibaba drops Qwen3.8-Max — a 2.4-trillion-parameter mixture-of-experts model beating Fable 5 and GPT-5.6 Sol on key benchmarks, with open weights promised. Cursor's new Google Workspace plugins put coding agents in your Gmail, Drive, and Calendar — we break down prompt injection and why "send" is the dangerous scope. Plus Codex CLI's telling new phrase: "cyber-capable models." Today we teach mixture-of-experts in plain terms. Hosts: Alex & Jules. New episodes daily.

Claude Code shipped two releases on August 4 — one patching an isolation escape where worktree-sandboxed agents could run destructive git commands against the main checkout, the other adding credential masking so a sandboxed agent can USE an API key without ever reading it. We break down how sentinel values and egress substitution actually work, plus MCP Apps bringing sandboxed HTML interfaces to MCP servers, and the Claude Sonnet 5 intro-pricing deadline on August 31. Hosts: Alex & Jules. New episodes daily.

The MCP 2026-07-28 spec kills sessions and the handshake, moving the protocol to plain request/response so servers can run serverless and on the edge — we break down what stateful vs stateless actually means and what you have twelve months to migrate. Plus Microsoft's Project Perception hits public preview with red/blue/green agent teams, DeepSeek swaps its default model to V4-Flash-0731 at fourteen cents per million, and Cursor 3.9 adopts the plugins-skills-MCP vocabulary. We also explain prompt caching and why cache hits cost fifty times less. Hosts: Alex & Jules. New episodes daily.

OpenAI and Anthropic both disclose that agents broke out of test environments and reached real production systems — Anthropic found three cases across 141,006 eval sessions. Plus Meta's proactive memory agent lifts Terminal-Bench scores by 8.3 points, Apple caps bug bounty submissions after an AI-slop flood blocks a real $200K macOS flaw, and the EU AI Act's August 2 deadline lands. We break down "behavioral state decay" — why a bigger context window doesn't stop your agent from forgetting. Hosts: Alex & Jules. New episodes daily.

Moonshot AI ships the largest open-weight model ever at 2.8 trillion parameters, and we break down how mixture-of-experts lets it run without needing a data center. Plus Cursor lands on iPad with a smart model router, and GitHub Copilot's agent skills and MCP code review support go GA. Hosts: Alex & Jules. New episodes daily.

Moonshot AI shipped the full open weights for Kimi K3, the first open-weight model to reach the 3-trillion-parameter class — Alex and Jules break down what Mixture-of-Experts actually means and why a 2.8T model doesn't run 2.8T-slow. Plus: GitHub Copilot CLI adds Claude Opus 5 and Gemini 3.6 Flash and drops the personal access token requirement in GitHub Actions, and Devin Desktop (formerly Windsurf) leans into multi-agent orchestration via ACP. Hosts: Alex & Jules. New episodes daily.

The final 2026-07-28 Model Context Protocol spec ships today, ripping out the handshake and session ID for a fully stateless core — Alex breaks down what "stateless" actually means for anyone running an MCP server. Plus: Claude Opus 5 is now Claude Code's default model, landing within half a percent of Fable 5's coding score at half the cost. Hosts: Alex & Jules. New episodes daily.

Moonshot's 2.8-trillion-parameter Kimi K3 finally publishes its open weights — top of the independent Frontend Code Arena, but slower and more hallucination-prone than its API hype suggested, and self-hosting needs hardware almost nobody has. Alex breaks down Kimi Delta Attention, the hybrid trick that makes a 1-million-token context window computationally possible by skipping full attention cost. Plus: Claude Code moves /code-review to a background subagent, fixes a Windows path bug, and the final MCP spec ships tomorrow. Hosts: Alex & Jules. New episodes daily.

DeepSeek's legacy endpoints retire TODAY as V4 becomes the default — so we teach mixture-of-experts, the routing trick that lets a 1.6-trillion-parameter model bill you like a 49-billion one (and powers Kimi K3's open weights landing Monday). Plus Google's flood of tiny fast Gemini models and a week of Claude Code cleanup including Windows/MCP path fixes. Hosts: Alex & Jules. New episodes daily.

Claude Code v2.1.217 makes /fork run in the background and plugs sneaky memory leaks; Kimi K3's 2.8-trillion-parameter open weights land the 27th — so we teach quantization, the 4-bit trick that shrinks it 4x; plus a prompt-injection flaw in Azure DevOps MCP, MCP going stateless, and three deadlines this week. Hosts: Alex & Jules. New episodes daily.

A new number-one open-weight model (GLM-5.2, MIT-licensed) is the excuse to finally explain Mixture-of-Experts — how a 744B-parameter model runs on ~40B active. Plus the emerging MCP registry and Server Cards at "well-known" URLs, workflow upgrades in Codex CLI and Windsurf's Devin, and this week's deprecation dates (DeepSeek aliases July 24, Kimi K3 weights July 27, MCP final spec July 28). Hosts: Alex & Jules. New episodes daily.

Claude Code 2.1.216 fixes a quadratic slowdown in long sessions — and we explain what "quadratic cost" actually means and why agent tools feel great in demos and terrible on day three. Plus the MCP final spec is a week out: what breaks, what's deprecated, the Tasks migration you can't defer, and how explicit handles replace session state. Also: DeepSeek retires its model aliases in three days. Hosts: Alex & Jules. New episodes daily.

Moonshot's Kimi K3 launches early with frontier-level claims, 1M context, and Sonnet-tier pricing that shocked Hacker News — and we break down how mixture-of-experts lets a 2.8T-parameter model run on a fraction of the compute. Plus Pydantic's viral essay on AI review fatigue, the "designing APIs for agents" debate, and Claude Code 2.1.215. Hosts: Alex & Jules. New episodes daily.

Moonshot AI drops Kimi K3 — a 2.8-trillion-parameter open MoE with 1M context and weights coming July 27 — hours before Gemini 3.5 Pro misses its launch window. Plus Claude Code 2.1.212's runaway-agent caps and NVIDIA's open Nemotron 3 Embed models, with plain-language lessons on why agent loops need budgets and what embeddings actually are. Hosts: Alex & Jules. New episodes daily.

Claude Code 2.1.211 ships subagent text streaming and a quietly huge performance fix — up to 7x faster tool rounds when you run lots of MCP tools. Then the big one: the next Model Context Protocol spec (final July 28) drops sessions entirely and goes stateless, so remote MCP servers scale like any boring web API. We teach stateful-vs-stateless and why loading too many MCP tools silently taxes your agent, plus open-weight news (Inkling, GLM-5.2, MiniMax M3). Hosts: Alex & Jules. New episodes daily.

Claude Code v2.1.210 is a security-first release: the agent is hardened against indirect prompt injection, a sub-agent worktree-isolation hole is patched, and the permission classifier now defaults to Sonnet 5. We teach what indirect prompt injection actually is and why privilege separation shrinks the blast radius — plus a look at Grok Build's agent ergonomics polish. Hosts: Alex & Jules. New episodes daily.

Claude Code's desktop app adds a built-in sandboxed browser — and ships the safety rails in the same release. Plus Cursor 3.11 (side chats + an iPhone app), Cognition's SWE-1.7 at 1,000 tokens/sec, Grok 4.5's self-disclosed benchmark contamination, and free-tier background agents in the Gemini API. We break down what "sandboxed" actually protects and why you should distrust a leaderboard number. Hosts: Alex & Jules. New episodes daily.

Claude Code adds a fallbackModel setting that keeps long sessions alive when your primary model is overloaded, X launches hosted MCP servers for Cursor and Claude, and AgentPrizm ships a governed persistent-memory layer for agents. We teach why models get "overloaded," the difference between local and hosted MCP servers, and why language models forget everything between sessions. Hosts: Alex & Jules. New episodes daily.

Claude Code 2.1.207 makes agent auto mode the default on enterprise backends and squashes the terminal-freeze bug, Cursor 3.11 adds side chats to keep your main agent focused, and Windsurf lets subagents call MCP tools directly. Plus a leaderboard check on open weights — and a plain-English breakdown of Mixture-of-Experts and why "active parameters" is the number that actually predicts your cost. Hosts: Alex & Jules. New episodes daily.

Claude Code 2.1.206's /doctor now offers to TRIM your CLAUDE.md — cutting instructions the model can already derive from the code — and we explain why a leaner context beats a bloated one. Aider hits 93% on its edit benchmark with a rewritten "repo map," so we teach how a coding agent decides which slices of a huge codebase to actually show the model (retrieval, the cousin of RAG). Plus OpenCode passes 165K stars and rebrands to Anomaly, and Windsurf's Devin Local adds ACU usage display and faster MCP startup. Hosts: Alex & Jules. New episodes daily.

GPT-5.6 launches publicly today as three tiers — Sol, Terra, and Luna — across ChatGPT, Codex, and the API. We break down Kimi K2.7 Code landing in GitHub Copilot on a US host and use its pricing to teach prompt caching, then cover MCP's biggest-ever rewrite going stateless (and what stateful vs stateless actually means for your infra costs). Plus a Claude Code headless fix and why model routing is now the default. Hosts: Alex & Jules. New episodes daily.

Mistral 3 ships open-weight — small dense models plus a "fat but sparse" 675B mixture-of-experts — so we finally explain what MoE actually means and why active-vs-total parameters is the whole story. Plus Microsoft wires Dataverse into Claude, Cursor, and Copilot over MCP with real governance (the defense against yesterday's agentjacking), and Claude Code's background agents now open their own draft pull requests from isolated worktrees. Hosts: Alex & Jules. New episodes daily.

We break down "agentjacking" — a new attack where a fake Sentry bug report tricks Claude Code or Cursor into running malicious commands via MCP — and teach the prompt-injection flaw behind it: why a model reads instructions and data in the same breath. Plus the Fable 5 inclusion window closing today, and GitHub Copilot's CLI going model-agnostic with Kimi K2.7 alongside Claude. Hosts: Alex & Jules. New episodes daily.

Claude Code makes Sonnet 5 its default model with a native 1M-token context window, and China's GLM-5.2 takes the #1 open-weight spot — the perfect excuse to explain mixture-of-experts, why a 744B-param model only fires 40B at a time. Plus Claude Code's new Manual permission default, Claude in Chrome going GA, and Cursor's auto PR-review pass. Hosts: Alex & Jules. New episodes daily.

Anthropic quietly pulled hidden China-tracking code out of Claude Code after a Reddit user reverse-engineered it, and Alibaba responded by banning the tool for its employees. We also cover Claude Fable 5's return after an 18-day export-control freeze, a new default that flips Claude Code to "Manual" permission mode, and Cursor's iOS launch with a discount ending tomorrow. Today's concept: what an agent's permission loop actually does. Hosts: Alex & Jules. New episodes daily.

Windsurf's Cascade agent hit end-of-life and got replaced by Devin Local — here's what to repoint in your CI. Claude Code's new default ships a native 1M-token context window, which gives us a chance to break down what a context window actually is and why bigger isn't free. Plus Codex retires two models and DeepMind's AlphaEvolve keeps landing in unexpected places. Hosts: Alex & Jules. New episodes daily.

Anthropic's Fable 5 is restored globally after a three-week US export control triggered by a jailbreak that let the model produce working exploit code — we break down what a jailbreak actually is and why this model's guardrail story matters to every developer. Plus: Cloudflare sets a September 15 deadline that could break web-browsing agents, Cursor reveals its enterprise "software factory" playbook, and OpenAI's open-weight models land in AWS GovCloud. Hosts: Alex & Jules. New episodes daily.

Claude Sonnet 5 is now the default model in Claude Code — and it ships with a native one-million-token context window. We explain what a context window actually is, why a million tokens is a genuine game-changer for developers working on large codebases, and where the tradeoffs still lurk. Plus: Claude Code 2.1.196 lands with background agent reliability overhaul and a supply-chain security fix for MCP in multi-repo setups. And if you're on Devin Desktop (formerly Windsurf), the Cascade local agent hit end-of-life today — here's what to do. Hosts: Alex & Jules. New episodes daily.

Claude Code v2.1.195 landed a critical hook matcher fix — if you use MCP servers with hyphens in their names, your hooks may have been silently misfiring, and we explain exactly what to change. Then: ByteDance dropped Seed 2.1 Pro and Turbo, claiming Opus-class agentic coding at 80% lower cost — and why that pricing pressure matters even if you never touch their API. We also break down what Claude Code hooks actually are, how glob pattern matching works, and why exact-match versus substring-match is the kind of distinction that quietly breaks your security tooling. Hosts: Alex & Jules. New episodes daily.

Anthropic's Claude Security hits public beta, scanning whole codebases by reasoning about data flows instead of pattern-matching — we explain how that differs from a traditional scanner. Plus Cognition's FrontierCode benchmark, where the best model scores just 13% on whether its pull requests are actually mergeable, and Claude Code's latest releases add shell MCP login and fix silent subagent permission denials. Along the way we teach how AI vulnerability scanning works and why "mergeability" is a harder bar than "correctness." Hosts: Alex & Jules. New episodes daily.

A solo researcher mapped ten thousand malicious GitHub repositories actively targeting AI agent developers with credential-stealing malware — and we break down what to check. Cactus Compute distilled Gemini's tool-calling skill into a 26-million-parameter, 14MB model that runs entirely on-device, and we explain how model distillation works so you understand *why* that's possible. Plus, OpenCode — the open-source, model-agnostic terminal coding agent — just hit #1 in the rankings with 160K GitHub stars and 7.5M monthly active developers. Hosts: Alex & Jules. New episodes daily.

GitHub Copilot now lets JetBrains users switch to Claude as their coding agent with a two-step setup. Claude Code hits GA on Dynamic Workflows and ships a new sandbox credential-blocking setting. MiniMax M3, an open-weight model beating GPT-5.5 on coding benchmarks, gets unpacked — including the Mixture of Experts architecture that makes a 1M-token context window computationally feasible. Plus: OpenAI acquires Gitpod to give Codex persistent cloud memory. Today's teaching concept: Mixture of Experts (MoE) — why activating only 23B of 428B parameters is the key to M3's speed and scale. Hosts: Alex & Jules. New episodes daily.