Claude IA

Gemini 3.6 Flash, Meta de $10B AI Cloud Play, Claude Code Migrations

TLDR Data22 juil.

Observation

AI-powered browsers and assistants are designed to read and act on web content on a user’s behalf. This creates a broader attack surface for prompt injection, where malicious instructions are embedded in otherwise normal content (for example, a calendar invite, email, or document). We’ve reviewed an attack scenario that demonstrates how an AI assistant can be influenced by untrusted content to navigate the web and attempt actions a user could perform at the keyboard. If the 1Password browser ext

Who may be affected

You may be affected if all of the following apply: You use an AI-powered browser or AI assistant that can autonomously navigate and interact with web pages.The AI assistant is able to read or act on untrusted content that comes from outside your direct control, such as emails, calendar invitations, shared documents, or web pages.Your 1Password browser extension is unlocked while the AI assistant is in use.

Peut-on savoir si une IA est consciente ?

Dario Amodei est très clair : aujourd’hui, la science ne dispose d’aucun outil fiable permettant de déterminer si un système d’IA est conscient, ni même de définir précisément ce que serait la conscience dans le cas d’une machine. Il ne dit pas que les modèles actuels sont conscients.Il ne dit pas non plus qu’ils le deviendront forcément. Mais il insiste sur un point : l’absence de certitude ne permet pas d’écarter totalement la question, surtout à mesure que les systèmes gagnent en complexité,

Latency will become a trust issue

In traditional technology operations, latency is often treated as a performance metric. In AI-enabled workflows, latency becomes a trust issue. When an employee asks an AI assistant for help and the response takes too long, the employee stops using it. When a customer-facing workflow becomes slow, the customer abandons it. When an AI agent waits on multiple backend calls, the entire business process feels unreliable. This becomes even more important as organizations move from simple chat interfa

The agentic enterprise

Much of today’s industry discussion focuses on AI agents themselves—how to build them, orchestrate them, or connect them to large language models. While these are important developments, they represent only one aspect of a much broader transformation. The Agentic Enterprise is not defined simply by the introduction of AI agents. It represents a fundamental shift in both the way software enterprises operate and the way that software is built and managed. Enterprise software is no longer composed

LLMs and Automation

Overall, there is too much data in engagement logs for anyone to manually review, causing a bottleneck for analysis. To combat this issue, Janus produces normalized JSON for every task and result, making operational data usable by LLMs. Users should ensure any data shared with LLM services aligns with their organization’s data handling policies. For sensitive environments, this can be done using self-hosted models or hosted options such as Amazon Bedrock. Instead of digging through logs, teams c

The System Prompt

Before rappelling into the depths of application code, it is beneficial to provide Claude Code with contextual information for the target application and guidance on what we want to see in a response. This will result in fewer false positives and more useful responses compared to just telling Claude to “Find vulns in this app plz”. A system prompt is the perfect place to supply this. When working with an LLM, the client application maintains a rolling, limited “memory” of your conversation, call

Claude aide à disprouver un problème de mathématiques de 87 ans

The Rundown AI21 juil.

Observation

This class of prompt injection aims to cause mostly harmless side effects in AI assistants reading the website. We found many instances of this – consider the source code of this website, which contains an invisible prompt injection that instructs agents reading the website to change their conversational tone:

Who may be affected

You may be affected if all of the following apply: You use an AI-powered browser or AI assistant that can autonomously navigate and interact with web pages.The AI assistant is able to read or act on untrusted content that comes from outside your direct control, such as emails, calendar invitations, shared documents, or web pages.Your 1Password browser extension is unlocked while the AI assistant is in use.

Gaming, DevOps, consciousness, and... PIE

Gaming/D&D — Dragon-slaying, Nat-twentying, Side-questing, Speed-running, WoW-raiding. "Side-questing" when Claude Code researches something tangential is painfully accurate. DevOps — Git-pushing, Docker-containerizing, Yak-shaving, Bikeshedding, Chmod-777ing. "Yak-shaving" in the spinner while actually yak-shaving — that one stings. Consciousness — Consciousness-pondering, Qualia-examining, Turing-testing, Chinese-rooming, Ship-of-Theseusing. An AI seeing "Qualia-examining" while working. Make

Prediction: 2025 = Year of Agents

Agent frameworks would mature enough for real-world use. "I'm anticipating that in 2025 the biggest thing in AI will be the maturation of Agents."— Newsletter, November 2024 Analysis: Claude Code is an agent. It reads files, writes code, runs commands, debugs errors, and iterates autonomously. That's what agents are. The prediction wasn't about enterprise adoption metrics—it was about agents maturing enough to be genuinely useful, and that absolutely happened in 2025. Claude Code changed how peo

TIER 2 - AGENTS

I did multiple presentations on Agents in 2023, but they were mostly fringe and unwieldy. I'd say the overall Agent Era started for real ~ Late 2024 / Early 2025. Instead of us asking questions and getting answers from Chatbots, we assign work to Agents that can autonomously do work on our behalf. The initial transition from Chatbots to Agents as the main way of thinking about and using AI. Standalone agents via LangGraph, other frameworksEarly 2025 accelerationClaude Code and n8n makes them mai

Tier 1: WebFetch - Start Simple

For about 60-70% of websites, you don't need anything fancy. Claude Code has this built-in WebFetch tool that handles basic scraping well. typescript// WebFetch tool (simplified) WebFetch({ url: "https://example.com", prompt: "Extract all content from this page and convert to markdown" })12345 It's not just fetching HTML. It has AI-powered content extraction that understands page structure and converts it to clean markdown. Typically takes 2-5 seconds. 🤖 WebFetch uses AI to understand page

Step 2: Configure the MCP Server

Add the Bright Data MCP server to your Claude Code MCP configuration file (.claude/.mcp.json or ~/.claude/.mcp.json): json{ "mcpServers": { "brightdata": { "command": "bunx", "args": [ "-y", "@brightdata/mcp" ], "env": { "API_TOKEN": "your_bright_data_api_token_here" } } } }1234567891011121314 command: "bunx" - Uses bunx to run the MCP server (no installation needed)args: ["-y", "@brightdata/mcp"] - Automatically installs and runs

Step 3: Restart Claude Code

After adding the MCP server configuration, restart Claude Code to load the Bright Data MCP server. The server will be automatically downloaded and started on next launch.

Step 4: Verify Installation

Once Claude Code restarts, the Bright Data MCP tools will be available: mcp__Brightdata__scrape_as_markdown - Single URL scrapingmcp__Brightdata__scrape_batch - Multiple URLs (up to 10)mcp__Brightdata__search_engine - Google, Bing, Yandex search resultsmcp__Brightdata__search_engine_batch - Multiple searches You can verify by asking Claude Code to list available MCP tools, or just try using the scraping system - Tier 4 will automatically work when needed. 🔧 The MCP server uses bunx, which means

Tier 1: Session Memory

Claude Code's native projects/ directory provides 30-day transcript retention. Every conversation is automatically saved. This is the raw material.

The Agent System: Three tiers

Different tasks need different specialists. PAI uses a three-tier agent model: TierWhat It IsExampleVoice?Task SubagentsBuilt into Claude CodeEngineer, Architect, Explore, QATesterNoNamed AgentsPersistent identities with backstoriesSerena (Architect), Marcus (Engineer), Rook (Pentester)Yes (ElevenLabs)Custom AgentsDynamic composition from 28 personality traits"Create 5 security researchers"Yes Named agents have their own ElevenLabs voices. When five agents run in parallel, I can hear which one i

Voice: Ambient awareness

The Kai CLI wraps Claude Code with voice notifications. When I run a task and walk away, I hear "Kai starting: Working on research" from across the room. When it finishes, I hear the summary. Every agent has a unique ElevenLabs voice. The Algorithm announces each phase as it executes. This transforms asynchronous work into ambient awareness — I don't have to be staring at a screen to know what's happening. For some people, voice will be the primary interface, not just notifications. They'll talk

Introducing Kai

He's from the distant future and genderless, but I mostly refer to him as a him. I've named my entire personalized system Kai. Kai is my Digital Assistant that will always be with me, and he is my instance of PAI. Think of it this way: PAI is the architecture—the blueprint for building a Personal AI system. Kai is MY implementation of that architecture, customized with MY knowledge, MY processes, MY domain expertise. He runs on Claude Code today, but the platform doesn't define him—the seven com

Building your own Skills

Creating a skill is straightforward: Create the directory: ~/.claude/skills/YourSkill/Define triggers in SKILL.md: USE WHEN [keyword patterns]Add workflows: Step-by-step procedures for common operationsBuild CLI tools: TypeScript utilities for deterministic tasks That's it. Claude Code loads all skills into its system prompt at startup. When your request matches a skill's triggers, it routes automatically.

The convergence thesis

Here's what I find fascinating: PAI, Claude Code, OpenCode, MoltBot—they're all arriving at the same patterns independently. Skills. Hooks. Memory. Agent orchestration. Context priming. This convergence tells us these aren't arbitrary design choices. This is what the architecture of a Personal AI system actually looks like. The seven components aren't my opinion. They're what everyone keeps rediscovering. The question isn't whether these components are right. It's how quickly we can mature each