Latest / The AI Practitioner Brief / Claude Duet, Shopify's Models, and eMudhra's Weak Key — October 22, 2026

Transcript

Claude Duet, Shopify's Models, and eMudhra's Weak Key — October 22, 2026 · The AI Practitioner Brief

• 0:05
  SPEAKER_03: Welcome to The AI Practitioner Brief,
• 0:07
  SPEAKER_03: the podcast from AIPractitioner.substack.com,
• 0:11
  SPEAKER_03: where I turn the latest AI and tech news into practical insights
• 0:15
  SPEAKER_03: for AI practitioners.
• 0:17
  SPEAKER_03: I'm Lina Faik,
• 0:18
  SPEAKER_03: an ML Engineer Consultant and Lecturer in Advanced Machine Learning
• 0:22
  SPEAKER_03: at Sorbonne University.
• 0:23
  SPEAKER_03: It's Thursday, October 22.
• 0:25
  SPEAKER_01: Thursday already.
• 0:27
  SPEAKER_01: Let's get into it.
• 0:28
  SPEAKER_03: With me is Jimmy, my co-host on The Brief, who asks the questions I'd skip.
• 0:32
  SPEAKER_01: Happy to skip straight to the news then.
• 0:34
  SPEAKER_03: Good, because today's news has a strange shape.
• 0:38
  SPEAKER_03: Anthropic just handed its new AI agent the keys to your inbox and your calendar.
• 0:43
  SPEAKER_03: In the same week, a security researcher found something much worse elsewhere.
• 0:48
  SPEAKER_03: A company trusted to vouch that websites are legitimate had handed out signing keys
• 0:53
  SPEAKER_03: so weak that one of them was the number four.
• 0:55
  SPEAKER_01: The number four, as in the digit.
• 0:58
  SPEAKER_03: As in the digit.
• 0:59
  SPEAKER_03: We'll get to it.
• 1:00
  SPEAKER_03: First the headlines, then three deep dives.
• 1:03
  SPEAKER_03: We'll cover Anthropik's new work agent,
• 1:05
  SPEAKER_03: then Shopify's bet on smaller models instead of bigger ones,
• 1:09
  SPEAKER_03: and the cybersecurity day bad enough that AI tools ended up on both sides of it.
• 1:14
  SPEAKER_01: That's a strange sentence.
• 1:15
  SPEAKER_01: Start with the headlines.
• 1:17
  SPEAKER_03: Recursive.
• 1:18
  SPEAKER_03: A startup building local coding models announced Flint.
• 1:22
  SPEAKER_03: It's a mixture of experts model,
• 1:24
  SPEAKER_03: meaning it only activates part of its parameters for any given request.
• 1:28
  SPEAKER_03: Flint is built to run coding tasks on your own laptop instead of in the cloud.
• 1:32
  SPEAKER_03: The company says it generates over 125 tokens per second on consumer hardware.
• 1:38
  SPEAKER_01: Is it actually out?
• 1:39
  SPEAKER_03: It's announced, not yet independently benchmarked that I could find.
• 1:43
  SPEAKER_03: Second, Anthropic released its first official versioned API since 2023.
• 1:49
  SPEAKER_03: Now developers can pin to a specific version instead of getting surprised
• 1:53
  SPEAKER_03: by changes underneath them.
• 1:54
  SPEAKER_01: That's overdue.
• 1:55
  SPEAKER_03: That's the AI side.
• 1:57
  SPEAKER_03: Elsewhere in tech today, Elon Musk is pushing to merge Tesla and SpaceX.
• 2:01
  SPEAKER_03: He's reportedly aiming to announce it by mid-2027,
• 2:05
  SPEAKER_03: before the current administration leaves office.
• 2:08
  SPEAKER_01: Why would the administration matter to a merger between two of his own companies?
• 2:12
  SPEAKER_03: Reporting doesn't spell that out, just that he wants it done on this political timeline.
• 2:16
  SPEAKER_03: Meanwhile, DuckDB,
• 2:18
  SPEAKER_03: the small,
• 2:18
  SPEAKER_03: fast database a lot of data teams already run for analytics, released version 2.0.
• 2:23
  SPEAKER_03: That's a long-term support release,
• 2:25
  SPEAKER_03: meaning it gets years of updates instead of just a few months.
• 2:29
  SPEAKER_03: It also adds a new mode that lets AI agents query the data directly.
• 2:33
  SPEAKER_01: And the last one?
• 2:34
  SPEAKER_03: One analysis argues most analytics teams won't survive the next five years.
• 2:39
  SPEAKER_03: That's because AI is starting to answer the ad hoc questions that used to be their job.
• 2:43
  SPEAKER_03: The argument is that teams need to move from fulfilling requests
• 2:47
  SPEAKER_03: to owning business judgment, or they get automated around.
• 2:50
  SPEAKER_01: Rough note to close headlines on.
• 2:52
  SPEAKER_01: Let's hear about the agent first.
• 2:54
  SPEAKER_03: Anthropic opened a research preview of Claude Duet yesterday, October 21st.
• 2:59
  SPEAKER_03: It's a standalone agent built on Opus 5.5, Anthropic's current top model.
• 3:04
  SPEAKER_03: Instead of waiting for you to type a prompt, it works in the background.
• 3:07
  SPEAKER_01: Background meaning it just does things on its own?
• 3:10
  SPEAKER_03: It checks your connected apps every few hours,
• 3:12
  SPEAKER_03: sends a short morning brief on what matters,
• 3:15
  SPEAKER_03: and keeps working on tasks without you watching it.
• 3:17
  SPEAKER_03: It reaches into over 6,000 apps through what Anthropic calls connectors,
• 3:22
  SPEAKER_03: basically pre-built integrations into tools like your email or calendar.
• 3:25
  SPEAKER_01: Yeah, what that sounds like exactly the kind of thing that goes wrong at 3 in the morning.
• 3:29
  SPEAKER_03: That's the part Anthropic built guardrails around.
• 3:32
  SPEAKER_03: Duet never sees your account passwords.
• 3:34
  SPEAKER_03: It needs your approval before sending a message,
• 3:36
  SPEAKER_03: signing a document, completing a payment, or deleting a file.
• 3:40
  SPEAKER_00: What if I approve a bad one by mistake?
• 3:43
  SPEAKER_03: There's a harder line underneath that one.
• 3:45
  SPEAKER_03: Duet won't trade securities or move money between your own accounts, even if you tell it to.
• 3:50
  SPEAKER_03: Anthropic built that in as a flat refusal, not a setting it can be talked out of.
• 3:54
  SPEAKER_01: Isn't this just a bigger version of Claude Code, the coding agent Anthropic already ships?
• 3:59
  SPEAKER_03: Related, but a different scope.
• 4:01
  SPEAKER_03: Claude Code works inside one project on a task you handed directly.
• 4:05
  SPEAKER_03: Duet is meant to run continuously across your whole work life,
• 4:08
  SPEAKER_03: checking multiple apps without being asked.
• 4:11
  SPEAKER_01: Is this the first agent built like this, or is everyone doing it?
• 4:14
  SPEAKER_03: It's a crowded race.
• 4:15
  SPEAKER_03: Meta has a similar proactive assistant called Muse, and OpenAI has one called Dots.
• 4:21
  SPEAKER_03: Duet's pitch is narrower.
• 4:22
  SPEAKER_03: It's built for work tasks specifically, not general life admin.
• 4:26
  SPEAKER_01: Why is everyone racing to ship this now?
• 4:28
  SPEAKER_03: Because the model doing the reasoning underneath finally got reliable enough
• 4:32
  SPEAKER_03: to work unsupervised for hours instead of minutes.
• 4:36
  SPEAKER_03: That's a recent shift, not something Opus 5.5 invented alone.
• 4:40
  SPEAKER_01: Has anyone actually used it long enough to know if the guardrails hold up?
• 4:43
  SPEAKER_03: Not publicly, no.
• 4:45
  SPEAKER_03: It launched yesterday.
• 4:46
  SPEAKER_03: Right now the only account of how it behaves is Anthropic's own.
• 4:49
  SPEAKER_01: Who even gets to try it?
• 4:51
  SPEAKER_03: It's gated behind a waitlist,
• 4:53
  SPEAKER_03: limited to pro and max subscribers in the United States through the web or
• 4:56
  SPEAKER_03: a standalone iOS app.
• 4:58
  SPEAKER_03: Nothing about it is generally available yet.
• 5:00
  SPEAKER_01: I'd bet that hard line around money moves the moment enough users complain it's too restrictive.
• 5:05
  SPEAKER_01: Companies usually loosen a safety default once the backlash is
• 5:08
  SPEAKER_01: about convenience instead of safety.
• 5:10
  SPEAKER_03: Maybe.
• 5:11
  SPEAKER_03: Anthropic hasn't signaled that,
• 5:13
  SPEAKER_03: but plenty of products have softened a rule like this under pressure before.
• 5:17
  SPEAKER_03: What's different is this isn't a filter on an output,
• 5:19
  SPEAKER_03: it's a constraint on what the agent is even allowed to attempt.
• 5:23
  SPEAKER_01: What's stopping another company from skipping that hard line entirely?
• 5:26
  SPEAKER_03: Nothing forces it.
• 5:27
  SPEAKER_03: It's a design choice, not a regulation.
• 5:30
  SPEAKER_03: The EU's AI Act, Europe's main AI law, doesn't require a specific list of banned actions.
• 5:36
  SPEAKER_03: It just requires a risk assessment.
• 5:38
  SPEAKER_01: So the safety margin here is just whichever lab felt cautious that week.
• 5:42
  SPEAKER_03: An agent that checks 6,000 apps every few hours and keeps working in
• 5:45
  SPEAKER_03: the background isn't free to operate.
• 5:48
  SPEAKER_03: That operating cost is exactly what Shopify spent this week trying to shrink.
• 5:52
  SPEAKER_02: Shopify, the e-commerce platform, what did they do?
• 5:55
  SPEAKER_03: Shopify published an engineering post on how it builds task-specific AI models
• 6:00
  SPEAKER_03: for Sidekick, its AI assistant for merchants.
• 6:03
  SPEAKER_03: The goal was to stop relying only on frontier models like GPT or Claude for every task.
• 6:08
  SPEAKER_01: Wait, didn't we just spend a whole block on a company shipping a huge new agent?
• 6:13
  SPEAKER_01: Now a company's going smaller on purpose?
• 6:15
  SPEAKER_03: That's exactly the contrast.
• 6:17
  SPEAKER_03: Shopify took one specific job,
• 6:19
  SPEAKER_03: A sub-agent that writes GraphQL queries,
• 6:21
  SPEAKER_03: which is the language Shopify's own systems use to pull and update stored data.
• 6:26
  SPEAKER_03: They trained a much smaller open-weight model to do only that.
• 6:29
  SPEAKER_01: Open-weight meaning?
• 6:30
  SPEAKER_03: The model's parameters are published,
• 6:32
  SPEAKER_03: so anyone can download and retrain it,
• 6:34
  SPEAKER_03: instead of only reaching it through an API like the Big Frontier models.
• 6:39
  SPEAKER_03: Shopify started with Nvidia's NemoTron 3.5 Lightning, a smaller open model.
• 6:44
  SPEAKER_03: They kept training it specifically on Shopify's own GraphQL examples
• 6:48
  SPEAKER_03: until it matched real merchant queries.
• 6:50
  SPEAKER_01: And a Frontier model couldn't already do that job well enough?
• 6:53
  SPEAKER_03: It could, just at a cost.
• 6:55
  SPEAKER_03: According to Shopify's own internal benchmark,
• 6:57
  SPEAKER_03: the fine-tuned model scored slightly higher on accuracy than
• 7:01
  SPEAKER_03: the Frontier model it replaced.
• 7:03
  SPEAKER_03: It also answered seven times faster and cost about 90% less to run.
• 7:07
  SPEAKER_01: What's actually making the smaller model as accurate?
• 7:10
  SPEAKER_01: Shouldn't a bigger model with broader training always win?
• 7:13
  SPEAKER_03: Not on a narrow job,
• 7:14
  SPEAKER_03: a model that only has to learn one repeated task doesn't need the breadth that makes
• 7:18
  SPEAKER_03: a general-purpose model so much bigger,
• 7:21
  SPEAKER_03: and bigger is what makes it slow and expensive.
• 7:23
  SPEAKER_01: How much does that save in real terms?
• 7:25
  SPEAKER_03: Across Shopify specialist models overall,
• 7:28
  SPEAKER_03: the company says self-hosting cut serving costs by 90 to 98% compared
• 7:33
  SPEAKER_03: to frontier pricing.
• 7:34
  SPEAKER_03: One deployment reportedly dropped from an estimated $27 million a year to about 1 million.
• 7:40
  SPEAKER_01: 27 million to 1 million is the kind of number that gets a CFO to rewrite the AI budget.
• 7:45
  SPEAKER_03: That's Shopify's own number.
• 7:47
  SPEAKER_03: I didn't find anyone outside the company who's run the same comparison.
• 7:51
  SPEAKER_03: But the mechanism behind it holds up on its own.
• 7:54
  SPEAKER_03: A narrower job needs a smaller model.
• 7:56
  SPEAKER_01: So this doesn't work for a five-person startup?
• 7:58
  SPEAKER_03: Not yet.
• 7:59
  SPEAKER_03: You need enough repeated volume on one narrow task to justify building
• 8:03
  SPEAKER_03: and maintaining your own fine-tuned model.
• 8:06
  SPEAKER_03: If the task only comes up occasionally,
• 8:08
  SPEAKER_03: the Frontier model's higher price per call doesn't matter.
• 8:10
  SPEAKER_01: What's the actual lesson for someone listening who isn't Shopify-sized?
• 8:14
  SPEAKER_03: Before reaching for the newest,
• 8:16
  SPEAKER_03: biggest model by default,
• 8:17
  SPEAKER_03: check whether the task you're solving is narrow and repeated enough that a smaller,
• 8:21
  SPEAKER_03: specialized model would do it just as well.
• 8:24
  SPEAKER_03: That's worth testing this week if you're running any AI feature at real volume.
• 8:28
  SPEAKER_01: That's your one action item for the episode then.
• 8:30
  SPEAKER_03: It is.
• 8:31
  SPEAKER_03: Everything else today is interesting.
• 8:33
  SPEAKER_03: This is the one thing worth actually doing.
• 8:35
  SPEAKER_01: Then the real moat isn't the model.
• 8:37
  SPEAKER_01: It's whether you can afford the team to build the smaller one.
• 8:40
  SPEAKER_03: Owning more of your own AI infrastructure,
• 8:42
  SPEAKER_03: like Shopify just did, also means owning everything that can go wrong with it.
• 8:46
  SPEAKER_03: Which is a good way into today's other story.
• 8:49
  SPEAKER_03: This was a genuinely bad day for internet security, and AI tools sat on both sides of it.
• 8:54
  SPEAKER_01: Both sides meaning what?
• 8:55
  SPEAKER_03: In one direction, an AI model helped find a dangerous bug before criminals did.
• 9:00
  SPEAKER_03: In the other,
• 9:01
  SPEAKER_03: a tool a lot of AI companies run every day got compromised
• 9:04
  SPEAKER_03: and started stealing credentials.
• 9:06
  SPEAKER_03: Start with the good direction.
• 9:08
  SPEAKER_03: OpenAI says its model, GPT 5.6 LSAIL, helped researchers find two critical flaws in PHP.
• 9:15
  SPEAKER_03: The programming language that reportedly runs about 70% of the Internet.
• 9:19
  SPEAKER_01: 70% is a huge share for one language.
• 9:22
  SPEAKER_03: It's an old one,
• 9:23
  SPEAKER_03: built into a huge share of existing websites,
• 9:25
  SPEAKER_03: including code nobody's touched in years.
• 9:28
  SPEAKER_03: The bug itself was two years old and had stumped human researchers the whole time.
• 9:32
  SPEAKER_01: How does an AI model actually find a bug like that?
• 9:35
  SPEAKER_01: What's it doing differently from a person reading the same code?
• 9:38
  SPEAKER_03: That part OpenAI hasn't published in technical detail.
• 9:41
  SPEAKER_03: What they have said is that researchers working under its Patch
• 9:44
  SPEAKER_03: the Planet security program used the model to go from the old bug to
• 9:48
  SPEAKER_03: a working exploit in about two weeks.
• 9:50
  SPEAKER_03: That's something that had stumped humans for far longer.
• 9:53
  SPEAKER_01: So it's faster at proving a known weak spot is exploitable,
• 9:56
  SPEAKER_01: not necessarily spotting something totally new.
• 9:59
  SPEAKER_03: That's the honest read.
• 10:00
  SPEAKER_03: And there's a sharper edge to it.
• 10:02
  SPEAKER_03: Once details started circulating,
• 10:04
  SPEAKER_03: Cloudflare,
• 10:05
  SPEAKER_03: which runs infrastructure for a huge share of websites,
• 10:07
  SPEAKER_03: reported that attackers got hold of the flaw.
• 10:10
  SPEAKER_03: They went after its own customers before every site could patch.
• 10:13
  SPEAKER_01: So the same speed that helps defenders also helps whoever's racing them.
• 10:17
  SPEAKER_03: Exactly the trade-off.
• 10:18
  SPEAKER_03: Now the other direction, where AI tooling was the one attacked.
• 10:22
  SPEAKER_03: Light LLM,
• 10:23
  SPEAKER_03: an open source gateway a lot of companies use to route requests
• 10:26
  SPEAKER_03: between different AI models, shipped a compromised update on October 21st.
• 10:30
  SPEAKER_01: Compromised how?
• 10:32
  SPEAKER_03: Someone got hold of a maintainer's GitHub account token.
• 10:35
  SPEAKER_03: They built a malicious version from a lookalike copy of the real project
• 10:39
  SPEAKER_03: and published it to PyPI,
• 10:40
  SPEAKER_03: Python's package index, and to Docker Hub, where container images are stored.
• 10:45
  SPEAKER_03: It was live for about two hours before anyone caught it.
• 10:48
  SPEAKER_01: Two hours still sounds like enough time to do damage.
• 10:50
  SPEAKER_03: It stole credentials, cloud keys, SSH keys, and API keys for tools like Cloud.
• 10:56
  SPEAKER_03: It also dropped code that could keep running on an hourly basis afterward.
• 11:00
  SPEAKER_03: Light LLM pulled the release and says it's contained.
• 11:03
  SPEAKER_01: So anyone who installed that exact version in that window has to assume they're compromised.
• 11:08
  SPEAKER_03: That's the right read,
• 11:09
  SPEAKER_03: and Light LLM told users to rotate any credentials that were on
• 11:12
  SPEAKER_03: a machine running it.
• 11:14
  SPEAKER_03: The third piece of today isn't AI at all, but I can't get past it.
• 11:18
  SPEAKER_03: A certificate authority called eMudra,
• 11:20
  SPEAKER_03: which other companies pay to vouch that a website is really who it says it is,
• 11:24
  SPEAKER_03: Had signing keys so weak that researchers cracked eight of them in under a second.
• 11:29
  SPEAKER_01: How weak is weak here?
• 11:30
  SPEAKER_03: One of the private keys was the number four, not a four-digit number.
• 11:34
  SPEAKER_03: The number four.
• 11:34
  SPEAKER_01: You're joking.
• 11:35
  SPEAKER_03: I'm not.
• 11:36
  SPEAKER_03: Imudra says a flaw in the ceremony that generated the keys back in April
• 11:40
  SPEAKER_03: of 2025 produced a second weak batch alongside the real ones.
• 11:44
  SPEAKER_03: That weak batch got published too.
• 11:46
  SPEAKER_03: Chrome, Firefox, Safari,
• 11:48
  SPEAKER_03: and Edge all blocked the affected certificates within days of
• 11:51
  SPEAKER_03: a researcher reporting it.
• 11:52
  SPEAKER_01: None of these three are the same failure.
• 11:54
  SPEAKER_01: What actually connects them?
• 11:56
  SPEAKER_03: They're not one failure, they're one day.
• 11:58
  SPEAKER_03: But they share a cause worth naming.
• 12:00
  SPEAKER_03: PHP's codebase,
• 12:01
  SPEAKER_03: Light LLM's release pipeline,
• 12:03
  SPEAKER_03: and eMudra's key ceremony had all been trusted for years without anyone rechecking
• 12:07
  SPEAKER_03: the assumption that they were still sound.
• 12:09
  SPEAKER_01: I'll take recheck your assumptions over AI is coming for your job any day.
• 12:13
  SPEAKER_03: Three stories, one thread today.
• 12:15
  SPEAKER_03: Anthropic drew a hard line around what its new agent can do with your money.
• 12:19
  SPEAKER_03: Shopify proved a smaller model can beat a frontier one on a narrow job for
• 12:23
  SPEAKER_03: a fraction of the cost.
• 12:25
  SPEAKER_03: And today's security mess showed what happens when nobody rechecks
• 12:29
  SPEAKER_03: a system that's been trusted for years.
• 12:31
  SPEAKER_01: I'll bet the eMudra story is the one still getting talked about in six months,
• 12:34
  SPEAKER_01: not because of the damage,
• 12:36
  SPEAKER_01: but because the key was the number four is just too good a line to let go.
• 12:40
  SPEAKER_03: That's Thursday.
• 12:41
  [music]