The AI Velocity Problem
Every professional currently operating in the AI space is struggling with a shared, often unspoken reality: the sheer impossibility of keeping up. We are trapped in a simultaneous loop of tracking raw capability shifts, categorising the beneficial, the compromised, and the problematic, while attempting to lift the base literacy of our organisations at a pace that stretches normal human capacity. This is our day job, it isn't theirs, yet.
The recent report released on Project Glasswing serves as a cold case in point. When Anthropic's restricted Claude Mythos model can autonomously surface more than 10,000 high- or critical-severity software vulnerabilities in under 50 days with a 90.6% true-positive accuracy rate, it focuses the corporate mind rapidly. I'm not going to write much more on this, the analysis is out there and demands attention.
None of this is to diminish the marvellous opportunities AI presents. The potential to unlock efficiency, solve complex structural challenges, and augment human capability across every sector is genuinely extraordinary and makes my heart hopeful. It is why I'm working in this field. But it is precisely because the technology is so potent that its downside risks require such cold-eyed evaluation.
AI geopolitics — the Sovereign Club and the Wild West
The immediate industry response to the AI-driven vulnerability exposure has been to lock down the capability. Mythos has been withheld from the public domain and ring-fenced within a closed, state-sanctioned defensive coalition.
This marks the official arrival of a two-tier AI ecosystem, bringing massive implications for the sovereign control of computing power. We are looking at an exclusive club where access is controlled by hyperscalers in tight coordination with the nation state. Who is in and who is out of this circle will dictate immediate defensive supremacy, although benefits will trickle down to those in alignment. As I've said before, I don't think hyperscalers will long have autonomy here. The defence, intelligence and global stability agenda is too high stakes for nations to agree to leave it in private hands. I see this as the new nuclear proliferation problem, but fragmented, difficult to control etc.
Yet, this ring-fencing is not sustainable, with the last but one models still pretty good at similar capabilities (see the Anthropic report of likely state sponsored AI enabled cyber attacks from Q4 last year on which I made similar points last year). And it is likely that centralisation of this level of capability is already gone. While the dominant global powers engage in an intellectual property arms race, the open-weight ecosystem ensures a rapid multiplication of alternative models. In the middle space, a highly transactional "cyber-attack-as-a-service" economy is already weaponising these already out there models against mid-sized and small organisations for straightforward financial gain.
The systemic vulnerability here is acute. I look at the current state of IT patching within our critical institutions, such as the NHS or equivalent public services globally, with genuine concern. These institutions are structurally unequipped to defend against an asymmetric, machine-speed offensive, but we need them to be.
The broader horizon
Cybersecurity is merely one frontier. The velocity and depth of change is already highlighting and exacerbating societal harms which are multiplying across multiple vectors:
Algorithmic manipulation for profit. Whichever side of the coin you look at, whether it is hyper-targeted radicalisation to extremist points of view or the deliberate monetisation of human insecurities and anxieties, the capacity of these models to exploit psychological vulnerabilities at scale is unprecedented. We've already seen this playing out with tragic consequences in our societies.
Token pricing and equity. As the physical compute cost of frontier inference rises, token pricing structures are shifting. How do we shield non-profits and vital public services from being priced out of essential AI-driven transformation?
Cognitive surrender. There is a distinct human risk in how we interact with these tools. We face a growing trend toward cognitive surrender, where AI is used poorly as a lazy substitute for thinking rather than a sharp, critical peer. The danger is a steady hollowing out of independent human analysis in favour of algorithmic convenience.
The institutional vacuum. The educational pipeline is fractured. If entry-level tasks are automated, our traditional early-career learning pathways disappear. Our schools and universities are teaching curriculum structures designed for a world that no longer exists, while our creative outputs are subjected to a wholesale data grab by hyperscalers.
The social media analogy
We've seen that where one organisation draws a moral line, the competitive dynamics of the market ensure another will immediately step into the vacuum. Government policy, education-as-protection, corporate governance as control and legal frameworks remain perpetually behind, leaving society vulnerable to the commercial incentives of the big hyperscalers, and bad actors. History offers a clear, sobering precedent here. We watched the unchecked expansion of social media drive unprecedented anxiety and psychological stressors in children and teenagers, monetising user distress through algorithmic rabbit holes. And yes, I'm looking at you, Silicon Valley. (And I'm forever amused by Instagram's confusion about whether I'm a knitting granny or a history obsessed centrist Dad or perhaps a data geek into DnD — I see your attempts to monetise through your biased algorithms and I one-up you by being an anomaly — also, that amigurumi is cute.) As a society, we are only now beginning to catch up with regulation here, decades too late.
If we take the pace of that social media transformation, multiply it by the exponential capability growth of AI, and accelerate it by an immensely fast adoption rate, the trajectory suggests we are in for a very bumpy ride. Which should probably be read as a classic British understatement.
It is a demanding outlook for a Saturday afternoon, sorry. I also don't have any solutions to all of this, but I'd love to discuss it further with those who are working on it.