According to a quotation from Anthropic's Frontier Red Team collected by Simon Willison and posted on 29 September 2026, the team evaluated several AI models on 100 randomly selected tasks from an internal Binary Exploitation benchmark. The team reported that the model referred to as GLM-5.3 developed 'full control flow hijacks' — a type of software exploit — in 4% of trials, while a model called Claude Mythos Preview did so in 6% of trials. The quoted text states that although GLM-5.3 performed below Claude Mythos Preview on this measure, 'a meaningful threshold has clearly been crossed,' because earlier models such as Claude Opus 4.6 and GLM-5.2 did not succeed in any of the tasks. The excerpt is drawn from an Anthropic Frontier Red Team piece titled 'GLM-5.3 and the spread of advanced cyber capabilities.' The details available come from a single quoted excerpt rather than the full underlying report, and the benchmark itself is described as internal to Anthropic. No further methodology, definitions or context were included in the material provided.
- GLM-5.3 produced full control flow hijacks in 4% of 100 benchmark trials; Claude Mythos Preview in 6%
- Earlier models Claude Opus 4.6 and GLM-5.2 succeeded in none of the tasks
- Quotation collected by Simon Willison, posted 29 September 2026, from an Anthropic Frontier Red Team report
What it means for you
AI security researchers at Anthropic say newer models are starting to write real software exploits — code that breaks into or takes over programs — where older models could not do it at all. This is an early-warning finding, not a product or a crisis. For most people and small businesses, it changes nothing about your day; it's a signal that AI's offensive security skills are creeping upward over time.
Who should care
People who work in cybersecurity, run IT for an organisation, or track how quickly AI capabilities are advancing.
Skip this if
You are a general user or small-business owner not responsible for security — this is a research data point with no action attached to it right now.
Sources: Simon Willison — read the original