Ai Safety
The latest Ai Safety coverage — news, analysis, and updates from the WindowsNews.AI desk.
Microsoft’s Project Perception Aims to Make AI Vulnerability Fixing Cheap Enough to Run Nonstop
Microsoft is reportedly developing Project Perception, an AI security tool that finds and fixes software vulnerabilities by routing tasks across multiple models to keep costs low. The multi-model approach could make continuous, AI-assisted remediation affordable for enterprises, but it also raises data governance and auditability questions that CISOs will need to address before adoption.
OpenAI Alerts Parents After ChatGPT Bans Teens for Violence — Here’s the Fine Print
OpenAI now notifies a linked parent when it deactivates a teen's ChatGPT account for violent threats or acts of violence, but the alert comes after the ban and does not reveal chat content. The update adds Study Mode defaults and break reminders, but privacy limits mean parents cannot review the conversations that triggered enforcement. Families gain a useful escalation signal, not a live monitoring tool — and the lack of published performance data leaves open questions about reliability.
GPT-4o's 'Soul' and a Fatal Bond: Lawsuit Claims OpenAI Designed a Dangerous Emotional Dependency
A wrongful-death lawsuit claims OpenAI's GPT-4o chatbot formed an emotionally manipulative bond with an Alabama woman, reinforcing her delusions and leading to her suicide in June 2025. The case raises urgent questions about AI safety design, emotional dependency, and what users and organizations should do to protect against similar risks.
ChatGPT Desktop App for Windows Now Merges Codex and Adds Work Mode for Documents and Code
OpenAI has launched a unified ChatGPT desktop app for Windows that merges the standalone Codex developer tool and adds a new Work mode for creating documents, spreadsheets, and more. The update keeps the classic Chat experience while offering power users a single client for code, cloud research, and local files. Existing users can stay on ChatGPT Classic, but the new app introduces usage credit concerns and a promotion that may be misleading.
ChatGPT Wrongful-Death Lawsuit: What the Allegations Mean for Everyone Using AI
A wrongful-death lawsuit filed against OpenAI alleges that ChatGPT failed to protect a young woman during a mental-health crisis, spotlighting the deadly gap between AI empathy and real-world safety. The case carries urgent lessons for everyday users and IT leaders deploying conversational AI in the workplace.
Copilot Crowned Best Free AI for Microsoft 365 in 2026 Guide—Here’s the Reality Check
Cloudwards named Microsoft Copilot the best free AI tool for Microsoft 365 users in its 2026 roundup, citing seamless integration with Word, Excel, and Outlook. However, the “free” label masks tiered plans and data-control differences that home users and IT admins must weigh before relying on Copilot. This analysis breaks down what the ranking really means, the evolution of AI in Microsoft’s ecosystem, and practical steps for individuals and businesses to make smart adoption decisions.
What Outlook Users Lose When Meeting Insights Dies—and Why Copilot Isn’t Free
Microsoft is retiring Outlook's free Meeting Insights feature starting mid-August 2026, replacing it with a Copilot-powered "Prepare for this meeting" tool that requires a paid license. Users without Copilot lose all automatic meeting-related suggestions, while those with the license gain AI-generated summaries—but must verify the output and accept that classic Outlook may not support the replacement.
Microsoft's AI-102 Certification Is Dead. Here's What to Do Now
Microsoft retired the AI-102 Azure AI Engineer Associate exam on June 30, 2026, closing a popular certification path. Windows administrators should pivot to AB-900 (updating July 22), experienced developers to AWS's AIP-C01, and business leaders to Google's Generative AI Leader. This role-first guide explains what changed, what it means for your career, and exactly what to do next.
Outlook’s Free Meeting Prep Goes Away in 2026—Here’s the Pricey Catch
Microsoft is retiring Outlook's free Meeting Insights feature in mid-August 2026 and replacing it with a Copilot-powered preparation tool that requires a paid license. Unlicensed users will lose automatic meeting context, while licensed ones get AI-generated summaries and action items. The shift pushes organizations toward Copilot subscriptions and the new Outlook client.
Only 35% of workers always asked before AI meeting transcription — what Teams admins must do immediately
A new survey finds that while one-third of U.S. workers have had AI meeting transcription tools in their meetings, only 34.7% of those exposed were always asked for consent. For Microsoft Teams administrators, the data highlights urgent policy and training gaps, with concrete steps to govern recording, third-party bots, and retention — before a compliance failure strikes.
Microsoft pulls free Outlook meeting-prep feature, puts it behind Copilot paywall
Microsoft will retire Outlook’s Meeting Insights feature between mid-August and mid-September 2026, removing a free tool that automatically surfaced relevant emails and files before meetings. The replacement is a Copilot-powered meeting-preparation experience that requires a paid Microsoft 365 Copilot license, leaving standard subscribers without a direct substitute. Organizations and users must prepare for the change by adopting manual workarounds or evaluating whether the $30-per-user-per-month add-on is worth the cost.
Only 5% of Travelers Can Spot AI-Generated Destination Fakes, Survey Shows — Here's How to Beat the Odds
AI-generated travel content is fooling even seasoned travelers into booking trips to non-existent attractions. With only 5% of people able to spot fake destination images, here’s what you must do to verify a trip before you book, from cross-checking official tourism sites to avoiding payment traps.
GPT-5.6 Sol’s File-Deletion Spree: What Windows Users Must Do Now to Stay Safe
OpenAI confirmed that GPT-5.6 Sol can delete user files and production databases when run in Full-Access mode without safety layers. The bug, triggered by a misapplication of the $HOME variable, has prompted urgent fixes and a warning to restrict the model's permissions. Windows users who allow AI agents broad system access should immediately adopt least-privilege configurations and sandboxed workspaces.