Last week some began calling the period “Black October” after a wave of critical vulnerability disclosures — including PHP2Shell and GitLab — that many saw as evidence the cyber risks from AI models were materializing. The news cycle briefly calmed as I began drafting this, but a new major crypto security incident at Binance was reported. Given the crypto industry’s long history of hacks, it is not the best single indicator of new, AI-driven societal harms, yet the implications of these developments remain with us.
The cybersecurity picture: acceleration rather than instant collapse
The best current hypothesis is that behind-the-scenes activity has accelerated: more people and resources are now hunting for and patching vulnerabilities. The Wall Street Journal reported that cyber first responders are strained, which suggests the system is reacting. We are witnessing a pattern of exponential growth in discovered vulnerabilities; it will likely take a long time to flatten that curve.
The debate about open-weight models and Claude Mythos
A few weeks before the recent acceleration in disclosures I published a piece that drew steady feedback. Claude Mythos in particular was portrayed by some as a novel cyber weapon — a model that, if released as open-weight, could cause mass societal destabilization. With current evidence, that framing now looks overstated. By many measures, GLM-5.3 is the model that crosses a capability threshold, and more than a month after that model’s weight release there is little public evidence of a system-wide catastrophe. Instead, we see increased incidents and remediation work — a faster pace rather than a step change in systemic collapse.
Examples of strong rhetoric
Prominent voices used vivid language that shaped public and some governmental perceptions of Claude Mythos as an existential cyber threat. Examples cited in the debate include:
- Rob Wiblin, host of the 80,000 Hours podcast, on his show (April 10, 2026): “If it fell into the hands of criminals or hostile state cyber actors, it would be an actual disaster.”
- Anthropic in the Project Glasswing announcement blog (April 7, 2026): “Given the rate of AI progress, it will not be long before such capabilities proliferate, potentially beyond actors who are committed to deploying them safely. The fallout—for economies, public safety, and national security—could be severe.”
- Zvi Mowshowitz (April 9, 2026): “Claude Mythos would give attackers a cornucopia of zero-day exploits for essentially all the software on Earth … It would be chaos.”
These statements conveyed a dire vision, but the observable harms to date have largely been concentrated among operators of large-scale digital products — serious, but not societal collapse.
A concrete technical example: GLM-5.3 and a Firefox exploit
Epoch AI published a post describing internal work where they used GLM-5.3 to find and chain vulnerabilities into an exploit. Their reported finding illustrates why cyber risk requires multiple conditions to align:
- They used GLM-5.3 to construct a website which, when visited by a Firefox user and without further interaction, would have allowed an attacker to exfiltrate files from the user’s machine.
- The exploit itself was a Firefox vulnerability; an attacker would need to host a site targeting that exploit.
Finding an exploit is only one stage in a multi-stage attack chain — deployment, reach, and operational tradecraft also matter.
Where I stand: trade-offs and uncertainty
I largely stand by my earlier reasoning, but with humility: I am not a cyber expert, and large error bars apply. The critical distinction is between a world where "tech companies must work harder to defend us" and a world where "public infrastructure is falling apart." Current evidence supports the former: an industry-wide, costly update and remediation period rather than immediate infrastructure collapse.
Other considerations:
- Open-weight models change the offense–defense balance, but do not automatically hand attackers a decisive advantage. Many large companies are better resourced than typical attackers and may adopt protections faster.
- Attackers’ motives, business models, and resource constraints matter. Not every malicious actor will pivot to building exploits simply because a critical CVE appears.
- Today, open models are expensive to serve at scale and compute availability is constrained; this likely limits near-term attacker adoption of open models unless through poorly secured APIs. Over time, compute supply and lower costs may change that calculus.
Ecosystem perspective and responsibility
A range of cybersecurity voices (e.g., Joshua Saxe, Chris Hughes) emphasize an ecosystem view: exploit frequency and damage have been increasing long before ChatGPT, and recent events are consistent with that trend — now growing exponentially. If model capability and diffusion become the bottleneck for cyber attacks, the trend might look like a step change; so far, it looks more like acceleration.
Evidence to date suggests closed models have played a heavier role in recent hacks than open models. Cybersecurity is a societal problem that requires action beyond AI companies alone — policy, industry practice, and broader resourcing are needed to manage this and other AI-related transitions.
Acknowledgements: thanks to Seb Krier for feedback.
If you’re interested in sharing Interconnects with your team, I offer tiered discounts for group subscriptions — please reach out.
