September 29, 2026 — Meta wants your enterprise budget, minus the details

Host0:32

MongoDB's stock fell more than seventeen percent on Monday, and the reason is that its CEO left for Meta. That's from TechCrunch. Meta announced Meta Enterprise Platform, a new push to sell its AI to businesses. The first offering bundles Meta's models, agents, infrastructure and business tools. Named products include Muse, Meta Business Agent, Muse API and Muse Code. The man running it is CJ Desai. He was MongoDB's CEO and president. Now he's Meta's Chief Enterprise Platform Officer, reporting straight to Mark Zuckerberg. His résumé also includes Cloudflare and ServiceNow. MongoDB brought back former CEO Dev Ittycheria as interim chief. Meta's announcement describes a full-stack effort, but gives no deployment architecture, no pricing and no launch date. Desai said the goal is turning Meta's AI stack into products and services companies can deploy in their own businesses. He also said security and privacy are built in from the outset. That's a company claim, and there's no documentation yet to test it against. TechCrunch notes this builds on the Muse assistant launch earlier in the month. Enterprise buyers may wait for details on security, privacy, deployment and support before calling it a platform. What can a company actually buy first, and how does Meta handle data isolation, service levels and model choice? The announcement doesn't say.

2:12

Meta gave you a strategy, Anthropic a price that didn't move. Claude Sonnet five point five, out Monday, lists at two dollars per million input tokens and ten dollars per million output. Same as Sonnet five. The model ID is claude dash sonnet dash five dash five, an upgrade, not a rename. So where's the saving? Anthropic says it typically uses fewer tokens, so tasks can cost up to thirty percent less. It also says output runs more than thirty percent faster. Both are Anthropic's own claims. You get a one million token context window and up to one hundred twenty-eight thousand tokens of output. It's on Anthropic's platform, AWS, Google Cloud, Azure, and GitHub Copilot. On Terminal-Bench four point zero, which tests multi-step command line work, Anthropic reports seventy point six percent. Sonnet five scored ten point three. On CursorBench four point zero, it's fifty-five point five, against fifty-seven point eight for Opus five point five. Anthropic ran both. For callers, there's adaptive thinking with effort settings. The default is Medium in Claude Code and High on the Claude Platform. Migrating from Sonnet with thinking off? You'll need the new between tools setting. Limits. Risky cyber requests can fall back to Sonnet five, and some microbiology or virology requests may be flagged in error. Rate limits weren't in what I read. And Simon Willison's pelican prompt at Max effort spent one hundred twenty-eight thousand tokens, one dollar twenty-eight by his math, and never produced the SVG. Verdict. Try it today on everyday coding or a well-scoped document task. Pick it in the Claude Code model picker, rerun a bounded task you gave Sonnet five, and compare real token spend. Stay on Opus for open-ended work. Anthropic says Opus is stronger there.

4:18

Jevgrep's README reports two cost savings; the complete rerun reports a smaller one. It's a command line tool for coding agents. You ask a question about a repository, and the command, jg, hands back relevant files, reading leads and verbatim source excerpts. It walks the repository hierarchy, picks files from content previews, then returns useful source units with surrounding context. It's retrieval for the agent, not an answer generator. The project's own ten-task SWE-bench comparison had both runs solving eight of ten. Reported Sol cost fell from seven dollars sixty-two to five dollars forty-four, twenty-eight point six percent, before counting Jev's own cost. The later rerun, including Jev's cost, reports a twenty-five point eight percent saving. The catch. Those were tuned development tasks, not a holdout, so it doesn't prove savings on your repo. And searches send eligible source content to Jev through your chosen provider. To try it, you need Node twenty-two, macOS or Linux, and an API key from a supported gateway. Run `npm install -g @dzhng/jevgrep`, then ask jg a question about your project.

5:38

Jevgrep helps an agent read your code. This next one lets an agent spend your money. Shopify's developer changelog says browser agents can now read and update checkouts using WebMCP tools. They're get checkout, update checkout and complete checkout, plus navigate to storefront. The agent lists the tools registered on the checkout page, then calls one. They act on the buyer's active checkout in the buyer's browser session, sharing state with the checkout UI. No new API, no merchant configuration. The buyer stays in control. Shopify's docs say the buyer must confirm before complete checkout places the order. Things like three D Secure hand control back to them. Shopify also warns agents to treat text in tool responses as data, not instructions, because it could carry prompt injection. Now the trap. In Chrome one fifty-three, passing a JavaScript object to execute tool fails to parse. You have to pass a JSON string, and Chrome planned to accept objects in one fifty-five. It's unclear how merchants or buyers audit or reverse an agent's edits. If you build on it, keep that string workaround in one place. Chrome one fifty-five is when you get to delete it.

7:02

GitHub Security Lab says an Android app with no permissions could silently swap OsmAnd's map tiles for ones hosted by an attacker. Ordinary map browsing becomes a location leak. The cause was an exported activity that took attacker-controlled intent extras. Those could change map settings and send loaded tile coordinates to the attacker's server. It's one of twenty-four Android vulnerabilities that GitHub's open-source taskflows had found and reported by the time of the post. The design matters more than the count. Instead of one giant find-bugs prompt, the workflow first identifies attacker-controlled mobile entry points. Then it applies prompts for mobile bug classes, like confused deputy and insecure broadcasts. Results land in SQLite. Rows marked likely vulnerable are leads, not confirmed bugs. The limit is in GitHub's own section heading. LLMs are good at finding vulnerabilities but struggle at estimating severity. GitHub warns of false positives and wants a mobile-savvy human to review. The post doesn't say how many of the twenty-four were confirmed and fixed. To try it, you need a GitHub Copilot license and premium model requests. Open a Codespace on the taskflows project and run the mobile audit script with your org and repo name. On a medium repository, expect an hour or two.

8:35

GitHub wants a human checking its agent's work. Manus is giving its agents a wallet. Manus says each agent in its new Cue app has its own email, phone number, wallet and computer. That's a lot of equipment to issue before a first performance review. Cue is a standalone personal-agent app for phone and desktop, built on Manus infrastructure. Agents can hand work to one another, and Manus says the user sets direction and makes the final call. It arrived with Manus two point zero, which Manus calls a new architecture, not a routine update. The Cascade agent harness, in one tested configuration, used twenty-three point two percent fewer tokens, finished twenty-eight point two percent faster, and cost thirty-two percent less. Those are Manus's own numbers. There are also Automations that start workflows from new email, calendar events, Slack messages or Notion updates, plus Cloud Computer environments. The announcement doesn't explain how an agent's spending budget is isolated, audited or revoked if it does something unexpected. Privacy-minded users will want that first. To try Cue, it's early access and free with the invite code MEETCUE. Access is limited and first-come, first-served, and iOS was pending App Store review.

10:04

Before you hand any agent a wallet, update your phone. iOS and iPadOS twenty-six point seven point one, macOS Tahoe twenty-six point seven point one, and macOS Sequoia fifteen point eight point one. Install these. Apple patched a CoreGraphics out-of-bounds write that a crafted file could potentially turn into code execution. The Hacker News quotes Apple saying it may have been exploited in an extremely sophisticated attack against specific targeted individuals on versions before iOS twenty-seven. Apple hasn't disclosed how many people were targeted, whether any attacks succeeded, or when exploitation began. Meta Product Security reported the bug, so that's Meta's second appearance today. SIMDjson five point zero. Files full of numbers parse eight to twenty-five percent faster than four point zero in the release team's tests. That's on GCC sixteen point one and one pinned Xeon core, so verify on your hardware. Static reflection is now officially supported, and compile-time key selectors pull named fields in one pass, whatever the key order. Positive integers from two to the sixty-fourth up to ten to the twentieth now come back as big integers, not malformed numbers, so test that edge. There's also an optional flag to parse NaN and Infinity, which JSON forbids.

11:33

AMD just agreed to buy a lab that wants to influence what those chips are built for. AMD and World Labs. AMD agreed to buy Fei-Fei Li's World Labs in an all-stock deal valued at about eight point two billion dollars, with closing expected by the end of twenty twenty-six, pending regulatory approval. World Labs builds models that generate and simulate interactive three D environments from text, images and video, and it has already been optimizing training and inference on AMD GPUs. Li will join AMD as executive vice president and chief scientist, reporting to Lisa Su. AMD's pitch is that studying how spatial and robotics models use compute will shape its hardware and software roadmaps. That's forward-looking, so watch for real software support.

12:29

Fast Boot. Three security stories, and the attackers barely had to knock. Carbonato botnet. ThreatDown says Carbonato hijacks Docker daemons exposed on port twenty-three seventy-five, then hands operators a Telegram-controlled agent. Its thirty-nine-line persona file ranks AI API keys as loot number one, which is better onboarding than most of us got. Artifactory flaws. InfoQ reports three actively exploited flaws that can enable authentication bypass on exposed, self-hosted Artifactory and potentially lead to persistent admin access. In some cases, attackers established admin access in under five minutes. Wiz notes that upgrading doesn't evict anyone already inside, so check for the guest who beat your standup. Supabase exposures. UpGuard found sixteen thousand three hundred twenty-six Supabase databases with readable tables. UpGuard says programmatically created tables—the way coding agents interact with Supabase—don't have row-level security enabled by default. One was a Philippines OTP service with over a hundred thousand text messages, protected mostly by optimism. One exposed port, five minutes, one unlocked table. That's not a break-in, that's a walk-through.

13:53

Today Meta launched an enterprise AI platform and hired MongoDB's CEO to run it, Anthropic shipped Sonnet five point five at the same price, jevgrep told your coding agent which files matter, Shopify opened checkout to browser agents, GitHub explained how its agent found twenty-four Android bugs, and Manus gave agents a wallet. Try this today. Run one bounded coding task on Sonnet five point five and Sonnet five, and compare the token bill. Links for every story are in the show notes. Follow The Daily Reboot in your podcast app, so tomorrow's episode finds you. I'm your host. Go patch your phone. See you tomorrow.