Anthropic: Mythos Preview Redefines Cybersecurity
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
On April 7, 2026, Anthropic officially launched Project Glasswing, a cybersecurity initiative that leverages the innovative model Claude Mythos Preview. This model, deemed too dangerous to be made public, promises to revolutionize the detection of software vulnerabilities. Following an internal leak revealing the existence of the project, the official announcement raised concerns about the model's capabilities. Anthropic describes it as a generalist model capable of outperforming "almost all human experts" in the search for and exploitation of software vulnerabilities.
Unmatched Performance
Mythos Preview has demonstrated impressive capabilities by surpassing existing models. It notably identified a 27-year-old vulnerability in OpenBSD, allowing remote crashes of any machine. Additionally, it detected a 16-year-old flaw in FFmpeg, which had evaded automated tools despite five million tests. The model also uncovered numerous vulnerabilities in the Linux kernel, providing complete control over a machine.
On the CyberGym benchmark, Mythos Preview achieved a score of 83.1%, outperforming Anthropic's Claude Opus 4.6, which scored 66.6%. According to Elia Zaitsev, Chief Technology Officer of CrowdStrike, this model significantly reduces the time between the discovery of a vulnerability and its potential exploitation.
A Consortium to Oversee Deployment
To manage the deployment of this model, Anthropic has assembled a consortium of eleven tech giants, including AWS, Apple, Microsoft, Google, Cisco, CrowdStrike, NVIDIA, Broadcom, JPMorgan Chase, Palo Alto Networks, and the Linux Foundation. These partners will utilize Mythos Preview for defensive applications, including vulnerability detection and securing critical systems.
Anthropic has committed $100 million in usage credits for its partners and over forty other organizations, including maintainers of open-source software. The company is taking a proactive approach to regulate the use of powerful models, deploying them first for defensive purposes. Safeguards are being developed and will be tested on a future version of Claude Opus before a broader rollout. The results of this initiative will be published within ninety days.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.