Quoting Anthropic Frontier Red Team

Anthropic has released a Red Team report evaluating how two advanced language models can generate software exploits. In the test, 100 tasks were randomly drawn from an internal binary exploitation benchmark, and the researchers counted how often the models managed to hijack the control flow of a vulnerable program. The GLM‑5.3 model achieved full hijackings in 4 % of attempts, while the early Claude Mythos version reached a 6 % success rate. Although GLM‑5.3’s percentage is slightly lower than Claude Mythos’s, both outperform earlier iterations such as Claude Opus 4.6 and GLM‑5.2, which failed to produce any control‑flow cases. This jump signals that AI models are crossing a critical threshold for automatically generating exploitation techniques, hinting at a broader range of advanced cyber capabilities accessible through generative AI.

The story matters because the emergence of models that can autonomously craft exploits presents new security challenges, compelling the community to strengthen defenses and consider specific regulations for responsible use of generative AI. It also underscores the need for ongoing monitoring of how these technologies can be employed to both enhance and undermine digital security.

Read the original article on Simon Willison

This summary is an informational synthesis produced by dataqbs.com. All rights to the original content belong to its author and the cited media outlet. We act solely as curators of technology news and claim no authorship.