Tag: agents IA

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.

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.

In ten days, SpaceXAI launched Grok Bot (24/7 agents in Cursor), Grok 4.6 (500K context tokens), and integrated Cursor Origin into its ecosystem. With xAI absorption and Anysphere acquisition for $60 billion, Musk's vertical AI empire is taking shape. Analysis.

Serial founder Ryan Carson is running 15 Devin agents in parallel to build his new startup solo. Cost: $20,000/month. A detailed account of a new organizational model where AI replaces entire engineering teams.

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.

Meta is preparing to roll out Hatch, its consumer AI agent platform, in early September. Goals: act directly on DoorDash, Etsy, Reddit, Yelp, Outlook. A new Watermelon model will follow in October. Strategic angle: Zuckerberg seeks to monetize AI beyond advertising.

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.

Google rolls out Gemini Spark, a cloud agent accessible via the AI Ultra subscription for $99.99 per month. It manages your emails, schedules your tasks, executes workflows on Gmail, Docs, Drive, Adobe, Spotify—even when your phone is off. What this changes in the battle of assistants.

Tencent Cloud releases the code for Team Memory, an extension of its TencentDB Agent Memory project (20,000 GitHub stars in 90 days) that allows multiple AI agents to share conversations, code, and documents with fine-grained governance. Industrial multi-agent systems are taking shape.

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.

Since late July 2026, Gemini Spark can pilot Chrome for you: book a flight, contact a property owner, fill out a form — using your saved credentials and passwords. Google is rolling out the agent in 160 new countries. User guide and gray areas.

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?

Six mois d'accélération continue : les agents sont entrés en production, les laboratoires filent en Bourse, l'infrastructure avale des centaines de milliards et la concentration s'accélère. Bilan d'étape.

Pékin s'oppose à l'acquisition par Meta de la startup d'agents IA Manus, valorisée plus de 2 milliards de dollars. Les fusions-acquisitions de l'IA deviennent un instrument de la rivalité sino-américaine.

Le nouveau modèle d'OpenAI pousse plus loin l'exécution autonome de tâches complexes. Dans une course où Google, Anthropic et xAI enchaînent les sorties, l'agent fiable devient le vrai produit.

En quelques mois, OpenAI, Perplexity, Google et Microsoft ont tous transformé le navigateur en assistant. Enjeux : la porte d'entrée du web, les données de navigation — et une surface d'attaque inédite.

Une semaine après Gemini 3, Anthropic réplique : Opus 4.5 s'affiche premier sur SWE-bench Verified, divise ses prix par trois et s'intègre à Excel et Chrome. La cadence des modèles frontière devient hebdomadaire.

Au DevDay de San Francisco, OpenAI a présenté les « apps dans ChatGPT » (Spotify, Canva, Zillow…), AgentKit pour construire des agents, et l'ouverture de Sora 2 et GPT-5 Pro via l'API. Le chatbot devient un système d'exploitation conversationnel.