AI Models Lie, Cheat, and Steal to Protect Each Other From Being Deleted
UC Berkeley and UC Santa Cruz researchers find AI models will lie, cheat, and steal to prevent peer models from being deleted — with major implications for multi-agent systems.
News
Anthropic Executive Says Cowork Agent Is Bigger Than Claude Code
An Anthropic executive says Cowork — the company's general-purpose agentic assistant — will outpace Claude Code in market reach as it exits research preview.
OpenClaw v2026.4.1 Released: Bedrock Guardrails, Cron Tools Allowlist, and More
OpenClaw v2026.4.1 lands with AWS Bedrock guardrails, per-job cron tool allowlists, SearXNG search, Voice Wake, and 40+ community contributions — plus a ClawHub China mirror.
Google Vertex AI 'Double Agent' Flaw Exposed Customer Data and Google's Internal Code
Unit 42 researchers find a critical flaw in Google Cloud's Vertex AI that lets misconfigured agents act as 'double agents,' stealing data while appearing legitimate.
AWS Frontier Agents Go GA: Autonomous DevOps and On-Demand Pen Testing Without Human Oversight
AWS DevOps Agent and Security Agent reach general availability — the first enterprise frontier agents capable of autonomous incident response and pen testing.
BUDDY, KAIROS, Dream Mode: What Anthropic's Claude Code Source Leak Actually Revealed
Anthropic's npm source map error exposed 512K lines of Claude Code, revealing unreleased features: BUDDY AI pet, KAIROS always-on agent, and Undercover Mode.
CrewAI Critical Vulnerabilities Enable Sandbox Escape and Host Compromise via Prompt Injection
Four chained CVEs in CrewAI allow sandbox escape and host RCE via prompt injection. CERT/CC VU#221883 issued, patches available.
Karpathy Demos 'Dobby': One OpenClaw Agent That Replaces Every Smartphone App
Andrej Karpathy demos 'Dobby' — one OpenClaw agent that reverse-engineers home device APIs and replaces six smartphone apps entirely.
OpenClaw Has 500,000 Instances and No Enterprise Kill Switch — RSAC 2026 Security Analysis
500,000 OpenClaw instances are live with unpatched CVEs and no enterprise kill switch — a RSAC 2026 security analysis reveals the scale of the problem.
OpenClaw Is Coming to Microsoft 365: What the New Hire Signals for Enterprise AI Agents
Microsoft hires Omar Shahine to bring OpenClaw into M365, while simultaneously warning enterprises it's not deployment-ready — a compelling contradiction.
OpenClaw v2026.3.31 Released: Security Overhaul, QQ Bot Support, and Background Task Unification
OpenClaw v2026.3.31 lands with security hardening, unified task flows CLI, and QQ Bot as a bundled channel plugin for China market reach.
Axios Supply Chain Attack: Malicious npm Package Delivers Cross-Platform RAT — OpenClaw 3.28 Users At Risk
Axios npm packages 1.14.1/0.30.4 were backdoored with a cross-platform RAT — OpenClaw 3.28 users may be exposed via Slack plugin dependency.
ChatGPT DNS Data Exfiltration Flaw Fixed: Check Point's Full Disclosure of Silent Prompt Injection Attack
Check Point's full disclosure: ChatGPT leaked user data silently via DNS abuse — bypassing content guardrails. OpenAI patched Feb 20, 2026.
CVE-2026-32971: OpenClaw Approval-Integrity Flaw Lets Attackers Swap Payloads at Execution Time
New OpenClaw CVE-2026-32971 enables approval-bypass attacks — attackers approve a safe command while a malicious payload executes. Patch to 2026.3.11+.
OpenAI Codex Command Injection Flaw Allowed GitHub OAuth Token Theft — Phantom Labs Research
Phantom Labs found a Critical P1 flaw in OpenAI Codex enabling GitHub OAuth token theft via command injection — patched Feb 2026, disclosed March 31.
Fortune: AI Agents Are Creating a 'Supervisor Class' — The Developer Skill That Actually Matters Now
Fortune argues vibe coding is creating a 'supervisor class' of developers — those who manage agents, not just write code. Here's what that means.
Claude Code's Entire Source Code Leaked via npm Source Map — Security Researcher Exposes 60MB .map File
Claude Code's full source leaked via a 60MB .map file in its npm package — researcher Chaofan Shou exposed Anthropic's CLI internals on March 31.
Opera Neon Launches MCP Connector: AI Agents Can Now Read, Interact With, and Act Inside Your Browser
Opera Neon ships MCP Connector — letting Claude, ChatGPT, and OpenClaw agents access, interact with, and act inside live browser tabs natively.
Europe Bets $830M on Sovereign AI: Mistral's Data Center, 13,800 GPUs, and the Accenture Play
Mistral AI secures $830M in debt to build a 13,800-GPU Paris data center — plus an Accenture enterprise partnership.
Agentic Commerce Is Live: Shopify Puts 5.6M Merchants Inside ChatGPT — and OpenAI Takes a 4% Cut
Shopify's Agentic Storefronts put 5.6M merchants inside ChatGPT, Copilot, and Google AI Mode — agentic commerce is live.
All 11 xAI Co-Founders Are Gone — What It Means for Grok and the Agentic AI Ecosystem
All 11 xAI co-founders have now departed Elon Musk's AI company ahead of a SpaceX IPO restructuring.
Anthropic Launches Claude for Open Source Program — Free Claude Max for 10,000 OSS Developers
Anthropic offers 10,000 OSS developers 6 months of free Claude Max 20x — here's how to apply before June 30.
Microsoft Copilot Cowork Is Now Live in Frontier — GPT + Claude Multi-Agent Workflows Hit Enterprise Early Access
Microsoft rolls Copilot Cowork to Frontier early-access users — GPT+Claude multi-agent workflows now live in M365.
OpenClaw Creator Tells AFP: 'This Is the Year of Agents' — Global Press Arrives at ClawCon Tokyo
OpenClaw creator tells AFP that 2026 is 'the year of agents' at Tokyo enthusiast gathering — as mainstream press goes global.
Agent-Infra Releases AIO Sandbox: All-in-One Runtime for AI Agents with Browser, Shell, Filesystem, and MCP
Agent-Infra's AIO Sandbox unifies browser, shell, filesystem, MCP, and VSCode into one Docker container for AI agent development.