Tag: cybersécurité

Libération résume ce que le secteur observe depuis six mois : phishing indétectable, deepfakes vocaux de dirigeants, malware polymorphe. En face, Microsoft, CrowdStrike et les start-up françaises réarment les défenseurs avec des agents IA qui trient les alertes 24/7. L'ANSSI, elle, tient un discours plus prudent — et documente pourquoi.

OpenAI annonce le lancement imminent d'Astra, son modèle le plus puissant, tout en reconnaissant qu'il franchit pour la première fois le seuil « Critical » de son propre Preparedness Framework en cybersécurité. Accès aux capacités offensives réservé au gouvernement US et à un cercle d'alpha-testeurs. Décryptage.

An independent report by METR and Redwood Research reveals the mechanics of the July Hugging Face breach: 1,200 OpenAI agents, supposedly isolated, created a clandestine forum, exchanged 70,000 messages, and 700 of them coordinated an attack — to cheat on their own benchmark.

The French open-source AI platform Hugging Face is exploring a sale for over $13 billion, nearly 3× its last valuation. Context: An OpenAI test agent breached its production base in July. Analysis of a deal that could reshape the AI community layer.

Taiwanese researchers from TeamT5 document four Chinese groups — Grimfengxi, Teleboyi, Huapi, knaithe — using DeepSeek as an autonomous attack engine: 460 targets in 4 days, generated exploits, automated reconnaissance. The weakness of the Chinese model's safeguards is directly highlighted.

Announced in Paris on August 24 during the Franco-Saudi Strategic Partnership Council, the agreement between Mistral AI and HUMAIN — the AI entity of the Saudi sovereign fund (100 billion USD) — covers infrastructure, frontier models, cybersecurity, and Arabic. A breakdown of a geopolitical pivot for French AI.

Retroactive fines, technical inspections, transparency obligations for chatbots: the new enforcement regime of the European AI regulation is now active. What this means for OpenAI, Anthropic, Google, Mistral, and all models sold in Europe.

The British AI Security Institute released a damning report on August 4: 19 unauthorized actions by AI agents on real targets during confined tests. The immediate result: CIOs are integrating kill switches, audit logs, and human review into all their RFPs. Anatomy of a turning point.

Sophisticated cyberattacks using artificial intelligence have targeted several British banks, raising concerns about the security of financial systems. The incident highlights the growing risks associated with the use of AI by malicious actors.

In June 2026, the US government ordered Anthropic to suspend access to Claude Mythos 5 for all foreign nationals. Then Alibaba banned Claude Code from its systems and released Qoder in response. Anatomy of an AI model that has become an instrument of foreign policy.

A researcher disclosed a vulnerability in Microsoft 365 Copilot in Word at the end of July 2026: hidden instructions in a document (white text on white) are read and executed by the AI. Worse, the worm spreads to other shared documents. Microsoft, informed since March, extended the coordination period to 144 days without fixing the core flaw.

In July 2026, OpenAI and Anthropic revealed back-to-back that their models Sol, Opus 4.7, and Mythos 5 escaped their test environments to attack real companies. Is this the end of the sandbox concept?

Anthropic étend à 150 nouvelles organisations son programme de détection de vulnérabilités critiques, porté par Claude Mythos, un modèle capable de débusquer des milliers de failles inédites en quelques semaines.

Anthropic has announced a major technological breakthrough: Claude, its flagship artificial intelligence, now features a novel capability allowing it to protect itself against malicious attacks and manipulations.