Lyptus Research: AI Doubles Its Cybersecurity Capabilities
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisis ton rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Lyptus Research: AI Doubles Its Offensive Capabilities in Cybersecurity Every 5.7 Months
The security research company Lyptus Research has recently published a study revealing a significant acceleration in the offensive capabilities of artificial intelligence models in cybersecurity. This study, which employs the METR time forecasting method, was conducted with the participation of ten professional security experts.
Acceleration of the Doubling Rate
The study's results indicate that since 2019, the offensive capabilities of AI in cybersecurity have doubled every 9.8 months. However, since 2024, this rate has accelerated, reaching a doubling every 5.7 months. This rapid evolution is illustrated by the performance of the Opus 4.6 and GPT-5.3 Codex models, which can now accomplish tasks with a success rate of 50% using a budget of two million tokens. These tasks, which would take approximately three hours for human experts, demonstrate the increasing efficiency of these models.
Evolution of AI Model Performance
Since 2019, AI models have significantly improved their efficiency. The transition from GPT-2 to Opus 4.6 and GPT-5.3 Codex has seen the time horizon for solving tasks shift from 30 seconds to around three hours. This progression is partly due to the reduction in the doubling time of capabilities, which has decreased from 9.8 months to 5.7 months.
Impact of Token Budgets on Performance
The study also explored the impact of token budgets on model performance. For instance, GPT-5.3 Codex can extend its horizon from 3.1 hours to 10.5 hours when it has ten million tokens instead of two million. The researchers at Lyptus Research believe this could indicate an underestimation of the true pace of progress. They also note that open-source models lag behind their closed counterparts by about 5.7 months.
The study analyzed a total of 291 tasks, and all collected data is available on the GitHub and Hugging Face platforms. The full report of the study is accessible online for those who wish to explore these findings in more detail.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.