After this morning’s (17 April 2026) news reports, most of you will know about the furore surrounding the non-release of Anthropic’s Claude Mythos AI model and its implications for cybersecurity (listen to this podcast to get some understanding). Cybersecurity will, however, not be the focus of this post; it will, of course, be metaphors.
The furore started with a post by Anthropic ten days ago, followed by lots of discussions online. I thought I’d have a look at what Anthropic had written, but soon lost track of the technical details and my eyes fell on footnote 1 which said: “As in the previous article, these exploits target a testing harness mimicking a Firefox 147 content process, without the browser’s process sandbox or other defense-in-depth mitigations.” There you go! Hmmmm! What should I make of that?
In the following I’ll first think back to what I was used to, namely metaphors in genomics. Then I delve into the metaphors that are used in the footnote and try to dissect them. After that I’ll try to reflect on the differences between genomic jargon and metaphors and AI jargon and metaphors. And finally, I try to fathom what this means for the fields of AI and genomics and for the public understanding and communication of science in these fields.
Metaphors in genomics and AI
I am used to metaphors. When I read something like “The human genome is the book of life, with twenty-three chapters called chromosomes” or “CRISPR–Cas9 is a pair of scissors for cutting DNA”, I understand that these metaphors, with all their faults, encapsulate quite well-defined mechanisms grounded in reality; for example, in the case of CRISPR scissors, something like a programmable bacterial immune system adapted for gene editing etc etc. I also understand that ‘book’ means sort of genome, chapter means sort of chromosome and, of course, letters means chemical components that make up the genetic ‘code’…..And when I see something like ‘genes are a blueprint’, ‘the Golgi apparatus is a warehouse’ and ‘lysosomes are vacuum cleaners’, I can imagine what they do, however partial that may be.
Now, what about the metaphors used in this footnote? We have ‘exploit’, ‘harness’ and ‘sandbox’. I stared long and hard at them but could not make out their meanings. So, I asked Claude and got the following:
Exploit = the attack/hack (more on that metaphor here)
Harness = the fake test setup (like a “rig” or dummy environment) (more on that metaphor here)
Sandbox = the security barrier that keeps dodgy web content away from the rest of your computer (more on this metaphor here)
When you zoom into these metaphors, they all seem to have their roots in software engineering and development but are now reused in the context of AI – which is not overly surprising.
Claude also helpfully translated the footnote for me: “As in the previous article, these attacks are aimed at a fake/simplified version of Firefox that developers use just for testing purposes. This test setup deliberately leaves out Firefox’s normal security protections (like the sandbox that isolates web content from the rest of your system), so it’s not a realistic target — it’s more of a controlled environment to demonstrate that a vulnerability exists, rather than a proof that your actual browser is easily hackable.”
Ok, I started to get it, especially after one more helpful paraphrase: “these hacks work on a stripped-down dummy version of Firefox, not the real thing with all its defences”.
Before I continue, I’ll highlight one more metaphor which jumped out at me after Melanie Mitchell sent me an article about Claude Mythos, warning me it might be a bit technical. It was. But the first word that stumped me, was, again, not a technical term, but the word ‘moat’: “the moat is the system, not the model”. It turns out that this word is quite common in economics and means something like competitive advantage, based not only on the AI model alone but the whole infrastructure of expertise etc. surrounding it. And I bet there are many other words/metaphors like that out there.
Grappling with AI metaphors
Why did I find it so difficult to get a grip on these metaphors? Why did I have a feeling that their metaphorical meanings evaporated in a puff of semantic smoke as soon as I read them? Is it just my lack of knowledge of AI, software engineering and/or economics? But I have an equal ignorance of genomics ….. What’s going on here, I wondered. And does this tell us anything about the fields of AI and genomics?
Genomics has a mainly scientific vocabulary: genome, gene, DNA, codon, allele, nucleotide, protein and thousands more. They all refer to something ‘real’, something out there in the world, however complex and not quite understood. This vocabulary includes metaphors like ‘code’ and ‘letter’ which have been integral to genomic theory, as well as metaphors like ‘book’ and ‘chapter’ which are more frequently used to talk about genomics than inside genomics itself.
By contrast the AI vocabulary is not primarily a scientific one. Alongside key terms like neural networks, transformer models, large language models, fine-tuning, retrieval-augmented generation, reinforcement learning from human feedback and so on, we have a mixture of metaphors derived from computer science and software engineering jargon (sandbox, exploit, stack, buffer), business strategy language (moat, leverage), as well as general-purpose metaphors (frontier, pipeline). So, a lay person must first learn the jargon of software engineering or economics to understand these metaphors at least a little bit.
Does this reflect something about the field? AI research sits at an unusual intersection of engineering, commerce, and science. The Anthropic post is written by a security team trying to communicate urgency to industry partners; the moat post is by a startup trying to position itself competitively. Both are simultaneously making technical arguments and business arguments, and the language reflects that. And there is another issue…
The opacity of AI
What AI systems (including Claude) do is fundamentally statistical: they learn patterns across vast quantities of text and data, and those patterns let them do things that look like reasoning. The extraordinary results in the Anthropic report, namely finding a 27-year-old bug (now that’s a metaphor even I understand) that had survived decades of expert review, are genuinely real. The bugs are real. The crashes are real. The ‘exploits’ work on real computers.
But to understand how and why AI systems do this is much more difficult than to understand why a drug binds to a receptor. Nobody designed the system to find the bugs. Instead, these capabilities emerged as a downstream consequence of general improvements in code, reasoning, and autonomy (so says Claude). That is quite different from how genomics works, where you can trace a causal chain from gene to protein to function….and from metaphor to function, too, as in ‘DNA encodes the genetic instructions for building, operating, and maintaining living organisms’…
In AI the outputs are real, but the mechanism is genuinely opaque, even to the people who built it. Although metaphors exist, a precise mechanistic vocabulary does not yet exist. This is why the metaphors listed above can feel a bit like smoke and mirrors to outsiders.
With genomics terms like ‘codon’ or ‘allele’ are opaque at first, but they are opaque in the way a foreign word is opaque. Once you learn what they mean, they stay meaning that thing. A codon is three nucleotides encoding an amino acid. The word is a stable label for a stable physical reality. And so is the metaphor of the codon as a ‘three-letter word’ in the genetic code. Some metaphors decay though as research progresses; see the decline in the fortunes of the ‘blueprint’ metaphor, for example…
The AI/tech words, like ‘sandbox’, ‘exploit’, ‘moat’, or ‘harness’, are doing something quite different. They are metaphors that have been repurposed from other domains of knowledge, but they have not yet got a purchase on something that we really understand as our knowledge of AI’s inner workings is still murky. This reminds me of some epigenetics metaphors which were ‘in search of a target’.
Implications for the public understanding of AI
Genomics has grounding because its objects (molecules, sequences, cells) have some physical existence independent of our descriptions. AI is grounded in mathematics and computation, which are also real, but the behaviour of large AI systems is emergent and poorly understood even by experts, so the language around it tends to borrow from wherever vivid metaphors are available: engineering, business, warfare. The fact that many metaphors are borrowed from economics, it seems, reflect the fact that the field is as much commercial as scientific.
What does that mean for public understanding of AI and for science communication? In the past, science communicators have been trained to be aware of some of the limitations and shortcomings of genomics metaphor (the gene is not JUST a blueprint!). Given the nature of AI metaphors, it might be good to also raise awareness of the limitations and shortcomings of some metaphors that are inherent to AI discourse. These limitations are different from those in other fields, such as genomics, for example.
As we have seen, the inner workings of AI are rather opaque and difficult to pin down. So, any metaphor you use will not be instantly illuminating, but can perhaps let a bit of light fall onto the murkiness of AI’s inner workings. And ‘better’ metaphors might not help here.
We have also seen that AI metaphors are borrowed from different ‘languages’ (software engineering, computer science, economics, and ordinary language). And here is the rub.
The software engineering metaphors (sandbox, exploit, buffer, stack) are borrowed but essentially neutral. They are tools of the trade, jargon that happens to use everyday words. They obscure meaning through technical specificity, but they are not hiding an agenda. Once explained, they are just descriptions of how things work.
The economics metaphors, on the other hand, are doing something quite different. ‘Moat’ doesn’t just describe a competitive advantage; it frames AI development as a competitive race in which the goal is to build defensible market position. The metaphors of ‘leverage’, ‘pipeline’, or ‘harness’ all carry an implicit assumption that AI is primarily an instrument of value extraction. These metaphors smuggle in a worldview based on wealth creation and power. In this they are like the genomic metaphors of ‘blueprint’ or ‘code’ which smuggle in worldviews of genetic determinism.
In AI, the economics framing quietly normalises the idea that AI development is a race with winners and losers, which shapes how the public thinks about regulation, risk, and who gets to decide. The race framing was, of course, not absent in genomics, far from it, but it described how the people involved in the science behave, whereas in AI it describes what the enterprise actually is (see ‘frontier model’).
To paraphrase Wittgenstein, with genomics, the metaphors used are temporary scaffolding — you use them to climb up to the window, the concept, then kick them away. By contrast, with AI, the metaphors are the window. There is no kicking them away because there’s nothing more literal waiting behind them, not yet anyway. And some of the metaphors don’t lead up to a concept at all but to commerce.
Acknowledgement: Thanks to Tania Duarte for talking to me about some of these metaphors and explaining their economic roots.
Image: Sandcastle with moat and seaweed, Swanage.
Making Science Public 
Leave a comment