A wave of critical vulnerability disclosures, dubbed ‘Black October,’ followed the announcement of flaws like PHP2Shell and GitLab, a moment the author considers as when cyber risks linked to AI models began to materialize. A new major security incident at Binance, the cryptocurrency platform, was then reported, with hackers suspected of stealing billions.
The author notes that the cryptocurrency sector has long experienced hacks, making it, in their view, poorly representative for measuring the transition from vulnerabilities to new harms for society. They believe the behind-the-scenes work has mainly accelerated, with more people and resources mobilized to fix flaws. The Wall Street Journal reports that first responders in cybersecurity are under strain, which the author interprets as a sign the system is working.
Would an open-weight model in April have been manageable?
The post revisits a point of debate. An open-weight model (open-weight) is a model whose parameters are published and can be downloaded by anyone. During the announcement of Claude Mythos, Anthropic’s model was presented as a new class of cyber weapon, whose dissemination into the wrong hands could have caused massive societal destabilization. For the author, this reading was mistaken: in their view, GLM-5.3 already crosses this capability threshold, and public elements have changed little since its weights were published over a month ago.
They even suggest that a leak of Claude Mythos in open access would likely have left the world ‘more or less’ intact. It would have led to a clear increase in cybersecurity incidents, but more an acceleration than a rupture. The author cites several statements from influential figures. Rob Wiblin, host of the 80,000 Hours podcast, said on April 10, 2026: ‘If this fell into the hands of criminals or hostile state actors, it would be a real disaster.’ Anthropic, in the post announcing Project Glasswing on April 7, 2026, wrote that ‘the fallout for economies, public safety, and national security could be severe.’ Zvi Mowshowitz, on April 9, 2026, spoke of ‘a cornucopia of zero-day exploits for practically all software on the planet.’
Flaws discovered, but a complex attack chain
An Epoch AI publication describes internal work using GLM-5.3 to spot vulnerabilities and chain them into a potential attack. The exploit found in Firefox illustrates, according to the author, the need for several conditions to align for harm to occur: an attacker must, among other things, set up a site exploiting this flaw, which remains possible but is just one step among others.
The author acknowledges having chosen the wrong word: the term ‘correct’ seems insufficient to describe the future of cybersecurity, which will involve a costly update to how we work with software. They believe open models contribute to the balance between attack and defense, and that companies would likely adopt these models faster than attackers, who are often less well-resourced. They assume most recent vulnerabilities were found by closed models, with a significant margin of error, and that open models remain costly to run. In their view, cyber research shows the rise in exploits and damage has been underway long before ChatGPT. They conclude that, at this stage, evidence remains very weak to implicate open models, and that closed models have played a more significant role in recent hacks.
