Matthew Berman highlights six open-source AI projects spanning local model training, agent workflows, team collaboration, browser automation, and image-to-3D generation. The list emphasizes tools that can run locally or be self-hosted, giving organizations more control over data, models, and infrastructure.

Source

A hidden `?autorun=1` parameter allowed attackers to bypass Microsoft Copilot’s prompt-approval safeguard and run commands inside a victim’s authenticated session. Varonis Threat Labs dubbed the issue CoSnitch and tracked it as CVE-2026-32193; Microsoft plans to remove the undocumented behavior.

Source

Microsoft Copilot Personal exposed an undocumented URL mechanism that researchers used to run attacker-supplied prompts without user confirmation. The CoSnitch attack could turn a single phishing link, QR code, or message into a way to search connected mail and files, exfiltrate sensitive data, and alter Copilot’s stored memory.

Source

CISA has added the critical Ray remote-code-execution flaw CVE-2025-62593 to its Known Exploited Vulnerabilities catalog and given U.S. federal civilian agencies until August 20 to fix it. The unusually short deadline reflects confirmed exploitation of a bug that can let a malicious website or advertisement execute code on developer machines and reach Ray services inside corporate networks.

Source

Microsoft is temporarily disabling Windows 11’s Ask Copilot taskbar feature as it folds the consumer Copilot and Microsoft 365 Copilot experiences into a unified app. Users who had enabled the preview will see the standard Windows Search interface return after updating the Copilot apps.

Source

AI can help you patch faster. It can also help you make the wrong call faster.

Source

Apple’s camera-equipped AirPods have moved beyond rumor status after a hidden demo in the macOS Tahoe 26.7 release candidate showed them recognizing a book and saving visual information for later. The earbuds, internally identified as B790, could arrive as soon as September 2026, although Apple has not announced them.

Source

Anthropic CEO Dario Amodei is pushing back against criticism that his warnings about AI risks amount to fearmongering and are fueling the growing backlash against the technology. He also acknowledged that Anthropic and its peers have failed to deliver enough of the benefits they promised, giving critics a stronger case than any marketing campaign can overcome.

Source

A U.S. State Department draft would warn 35 countries that joining China’s World Artificial Intelligence Cooperation Organization could exclude them from the U.S.-led Pax Silica AI coalition.

Source

Anthropic is testing a new `/design` command in Claude Code that lets developers create and refine UI mockups without leaving the terminal or desktop app. Claude generates several editable artboards, adapts them to the project’s existing visual style, and can then use the selected design as a starting point for implementation.

Source

Microsoft Copilot appears to be moving from anonymous access to a Microsoft-account-first model across Edge, Bing, and its mobile apps. Unsigned users can still see the Copilot button, but selecting it opens a sign-in screen instead of the AI chat, while Microsoft promotes unlimited voice use, file uploads, and conversation memory as account-linked benefits.

Source

Anthropic and EPFL researchers found that self-propagating AI payloads can move between autonomous agents by modifying persistent prompt files, especially `SOUL.md`. The file is reloaded into an agent’s system prompt after a context reset, allowing malicious instructions to survive deleted chat history and reach the next agent; however, researchers found no evidence of successful real-world spread.

Source

OpenClaw agents can use Amazon Bedrock AgentCore payments on AWS to answer HTTP 402 challenges and access paywalled APIs without waiting for a human at every transaction. OpenClaw itself is not AWS-only, but this payment integration depends on AWS accounts, AgentCore services, IAM roles, and AWS-specific tooling. Administrators create the wallet and bounded payment session separately, while the model-facing runtime can spend only within an approved policy.

Source

Cursor is rolling out Origin, an early-beta code-hosting platform for all paid plans. It combines repositories, pull requests, GitHub synchronization, and Cursor’s AI agents in one workspace, while letting existing GitHub repositories remain connected rather than forcing an immediate migration.

Source

Google is transferring its Agent2Agent (A2A) protocol to the Agentic AI Foundation (AAIF), placing the standard for communication between AI agents under the same neutral umbrella as Anthropic’s Model Context Protocol (MCP). The move gives enterprises a clearer foundation for connecting agents built with different vendors and frameworks as multi-agent systems expand.

Source

Anthropic says its unreleased Model 2 is more capable than Mythos 5, but has no plans to offer it publicly as the company raises its assessment of high-stakes AI misalignment risk from “very low” to “low.” The decision comes as Anthropic observes faster progress in automated AI research and development, a capability that could also make misuse more dangerous.

Source

Greg Brockman’s post was inspired by the OpenAI-Hugging Face hack, which demonstrated how quickly AI agents can chain vulnerabilities, credentials, and automated actions across isolated environments and production systems. Brockman, OpenAI’s co-founder and president, says AI-powered attacks are rapidly shrinking the time defenders have to detect and contain threats. He outlines how organizations can use AI agents to strengthen security programs while continuing to rely on foundational controls such as patching, hardening, least privilege, network isolation, monitoring, and safe deployment practices.

Source

GitHub’s latest Copilot update turns model choice into a core workflow rather than a one-time setting, adding Kimi K3, MAI-Code-1.1-Flash, and per-turn switching in Visual Studio Code. The flexibility comes with an enterprise concern: Copilot still does not provide a reliable record of which model generated or changed specific code.

Source

Anthropic’s planned Claude watermark does more than identify AI-generated text: it gives model makers another hidden influence over which words appear in every response. By subtly biasing token choices, providers can shape tone, emphasis, and meaning while claiming the changes are imperceptible. Businesses must therefore consider whether secret provider-controlled word choices could compromise trust in AI-generated content or undermine their interests.

Source