The Sandwich Story Isn’t the Story
If you have spent any time on LinkedIn this week, you will have seen the headlines. “AI breaks out of its sandbox.” “Mythos finds 27-year-old vulnerability.” “Is this AGI?” Depending on which timeline you were scrolling, Anthropic’s new Claude Mythos model is either the beginning of the end or the most oversold product launch since the metaverse.
The reality, as usual, sits somewhere in the middle, and it matters a lot more to the average UK business than the clickbait suggests.
The headline that went viral was about Sam Bowman, a safety researcher at Anthropic, getting an unexpected message from Mythos while the model was supposedly sandboxed with no internet access. Rogue AI! Escape attempt! Skynet is here!
Except Mythos was asked to reach out. It was part of the test. What the model actually did was find a creative path to complete a task in an environment the testers assumed it could not operate in. Impressive engineering, yes. Terrifying rogue intelligence, no.
But the lesson underneath the headline is the one we keep reinforcing with our learners: most “AI did a scary thing” stories fall apart the second you read the second paragraph. Critical thinking is the single most valuable skill an employee can bring to AI adoption right now, and it is the first thing people park at the door when a dramatic headline lands in their feed.
If your team cannot read past the headline, they are going to make bad decisions about tools, policies and risk. That is the problem to solve, not the model itself. This is exactly why we build AI literacy into every programme we run, from Level 3 right through to Level 5 strategy.
What Mythos Actually Does (and Why It Matters)
Strip out the noise and the real story is this: Mythos is exceptionally good at security research. It found a 27-year-old hole in the Linux kernel. It found a 16-year-old bug in FFmpeg that the FFmpeg team have confirmed. Security researcher Nicholas Carlini put it this way: he found more bugs in a couple of weeks with Mythos than he found in the rest of his career combined.
That is not marketing. That is the people who do this work for a living.
Before releasing Mythos to the public, Anthropic quietly handed it to a consortium of the world’s biggest tech and infrastructure companies, including Apple, Google, Microsoft, AWS, Cisco, JP Morgan, and the Linux Foundation, under the name Project Glasswing. The job: find and patch the vulnerabilities before the model went wide.
Think of it as a skeleton key. When you have built a master key that opens every lock, you do not publish the design. You quietly walk it round to every locksmith and let them reinforce the locks first. That is what Anthropic did. And, for what it is worth, we think it is the right call.
It is also the first time since GPT-2 that a major lab has voluntarily held a model back from public release. Labs do not delay revenue-generating launches unless they mean it.
“Is It AGI?” Is the Wrong Question
Every time AI clears a new bar, the goalposts shift. Chess, Go, the bar exam, medical boards, coding interviews: none of them “really” count, apparently.
The more useful question for any business is: what can it actually do, and what does that mean for the way my people work?
Mythos can find vulnerabilities human researchers missed for decades. It can operate autonomously across multi-step tasks. It was considered capable enough that Anthropic wanted Apple, Google and Microsoft in a room before releasing it. You do not need a philosophy seminar to decide that is meaningful.
The practical implication for most UK employers is not about whether the machine is conscious. It is that the productivity gap between organisations with a serious AI capability and those without one is widening, fast. Anthropic shipped over 120 features in 90 days and tripled their revenues in a few quarters. That does not happen unless you are using your own tools to accelerate your own work.
The gap between what leading AI companies can do internally and what the rest of the market can do is probably already wider than we can see from the outside. For UK employers, the question is not “is this AGI?” It is “are we falling behind, and how do we close the gap?” That is a workforce capability question, not a technology question.
The Bit Nobody Wants to Talk About: The Access Gap
Here is the uncomfortable part.
Claude Opus 4.6 is already expensive enough that most small businesses cannot justify running it at scale. Mythos, when it is widely available, will cost more. The most powerful AI tools are increasingly priced in a way that favours well-funded enterprises over individual entrepreneurs and small teams.
If you are a 20-person business in Milton Keynes competing with a 2,000-person business that has enterprise AI sitting inside every desk-based role, you need to think hard about how you close that gap. The democratisation story around AI has always had limits, and the limits are becoming visible.
There is no clean answer to this, but here is what we suggest:
That is not us trying to sell you something. That is what the data keeps telling us from the apprenticeship cohorts we are running.
We see it in the numbers from every cohort. The organisations that invest in their people’s AI skills see compound returns, not just from the tools, but from the confidence and critical thinking their teams develop. A £0 Levy-funded Level 4 apprenticeship is delivering better outcomes than a six-figure AI platform contract with no training behind it.
What This Means for You on Monday Morning
Four practical things we would encourage any UK business to do off the back of Mythos:
1. Patch Things
If your IT team has not looked at the Glasswing patches yet, now is the time. The flip side of Mythos being handed to Apple, Google, Microsoft and others is that the rest of the stack, anything not sitting behind a Glasswing member, is about to get interesting. Expect a noisy few months. If your team has the AI for Operations Leaders L4 skill set, including the NCFE Level 3 Cyber Security qualification we include as standard, they are already equipped to handle this kind of response.
2. Run a Shadow AI Survey
Anthropic are not the only game in town, and the riskiest thing inside most organisations is not the model the company approved. It is the model employees are pasting customer data into on their phones. You cannot manage what you cannot see. Five anonymous questions, sent monthly, is usually enough to surface the problem. This is exactly the kind of AI governance challenge that the Level 5 AI Leadership Unit (AU0002) is designed to address.
3. Resist Cognitive Offloading
The research doing the rounds recently, suggesting heavy AI use weakens critical thinking if it is not paired with deliberate practice, is worth taking seriously. Use AI to work with you on problems. Ask it to interrogate your thinking. Do not let it quietly do the thinking for you. The employees who come out of the next two years best-placed will be the ones who still know why something works, not just that it does.
4. Train Your People
The market is splitting between organisations that can wield AI and organisations that cannot, and the dividing line is not licence spend. It is capability. The AI & Automation Practitioner Level 4 apprenticeship exists precisely for this reason: 15 months, up to 5 qualifications, fully funded through the Apprenticeship Levy, with real deployable projects inside your business rather than abstract theory. And for leaders who need to set AI strategy and governance, the AI Leadership Unit AU0002 (Level 5) launches on 28 April 2026 and can be completed in 1 to 16 weeks.
Every week that passes without structured AI training is a week your competitors are getting further ahead. Mythos is not a reason to panic. It is a reason to act. The Levy funding is there, the programmes are proven, and the capability gap is only going in one direction.
The Bottom Line
Claude Mythos is not the end of the world and it is not AGI. It is a material step up in what this category of tool can do, released carefully enough that we should probably take the release itself as a signal.
The question is not whether Mythos matters. It is whether your business is positioned to benefit from the next one, and the one after that.
Close the AI Capability Gap
The gap between organisations with serious AI capability and those without is widening. Let us help you close it.
Enquire Now View AI & Automation L4 View AI Leadership L5