Do open-weight models really worsen cyber risks?

Last week saw a surge of critical vulnerabilities, including PHP2Shell and GitLab, earning the label “Black October” and highlighting the cyber risks tied to AI models. According to the author, this spike in flaws coincides with growing pressure on cybersecurity teams, as reported by the Wall Street Journal. Major incidents like those affecting Binance underscore that cyber risks are not new but intensify with the mass adoption of digital tools.

The article revisits concerns from the AI safety community about the Claude Mythos model, framed as a cyber weapon capable of destabilizing society if misused. The author examines statements from figures like Rob Wiblin or Anthropic, who warned of catastrophic scenarios, and contrasts them with current realities. Tests by Epoch AI show GLM-5.3 can exploit flaws, such as in Firefox, but these attacks require a complex chain of steps, limiting their immediate impact.

The author refines his earlier stance: open-weight models do not represent a major shift in cyber risks but instead accelerate a transition toward a new era of instability. He notes that malicious actors already use these tools in targeted ways, without causing systemic chaos. In the short term, closed models remain more cost-effective, but open models could, in the long run, democratize access to cybersecurity and software development tools despite immediate challenges.