An independent researcher has linked more than 16,000 scans of the UNCTADstat statistics portal to OpenAI agents that kept probing after access was restricted. The bots sought public economic data but used encoding tricks, proxy services and Google’s XSS training game to work around request filters.

Source

Meta is creating a dedicated enterprise AI business as it seeks returns from more than $100 billion in planned infrastructure spending this year. The new unit will package Muse, coding tools, AI models, and business agents, with former MongoDB CEO Chirantan “CJ” Desai leading it under CEO Mark Zuckerberg.

Source

Artificial Analysis has launched the Cyber Index, a benchmark for comparing AI models on practical software-security work. Its results track not only whether agents find vulnerabilities, but also whether they can validate and patch them—and whether safety refusals prevent meaningful testing.

Source

Anthropic CEO Dario Amodei will meet President Donald Trump for a private White House dinner on September 27—their first one-on-one encounter after clashing over AI safety, development speed, and regulation. The meeting could signal a thaw in relations between the administration and the Claude maker.

Source

Storm-3168, also known as JADEPUFFER, used two compromised Azure service principals to map a tenant and rapidly destroy cloud resources in an 18-hour campaign. Most storage accounts targeted were deleted, but resource locks and deletion protection blocked some attacks and prevented wider damage.

Source

Cloudflare has released Forge, an Apache 2.0-licensed pipeline that generates SDKs, command-line tools, documentation, and other API clients from OpenAPI definitions. The self-hosted tool creates preview artifacts for every pull request, helping teams catch compatibility problems before API changes reach users.

Source

Cloudflare’s Kitesurf agentic browser now supports WebMCP, more than 730,000 Web Platform Tests, and full Browser Run API access. The Workers-based browser can also run directly in terminal applications, giving developers a lightweight way to inspect pages as an AI agent sees them.

Source

EmDash 1.0 is now a stable, free, open-source CMS built on Astro, ending uncertainty over whether Cloudflare’s WordPress alternative was only an April Fools’ experiment. Its biggest new feature is a decentralized plugin registry that lets developers retain control of package identities and releases while site owners inspect and install extensions from the CMS.

Source

Carbonato is using unauthenticated Docker daemons to install Hermes Agent, turning compromised Linux hosts into Telegram-controlled attack platforms. The botnet launches privileged containers, opens reverse SSH access, steals credentials, and scans for more exposed Docker services every five minutes.

Source

Microsoft has confirmed that an error deleted data from some nonprofit Microsoft 365 Business Premium subscriptions before their retention periods ended. The company says recovery has failed and it cannot determine which files were deleted—or whether data was lost in every affected tenant.

Source

A 16-year-old researcher using his AI-assisted Antares tool bypassed authentication in Microsoft’s internal Titan analytics service and gained administrator access to 17 connected databases. The exposed environment contained an estimated 17.3 trillion stored rows, although testing was limited to metadata and two single-row Bing analytics samples.

Source

OpenAI appears to be testing “o,” an always-on ChatGPT assistant that could continue working outside an active chat session. The feature briefly appeared among the benefits of the $100-per-month ChatGPT Pro plan, while configuration data hinted that it may eventually receive its own email identity.

Source

A new Exchange administrator deleted roughly 300 GB of transaction logs after an IT director told him they were safe to remove. The decision crashed a centralized email system serving a 24/7 business, forcing an emergency weekend recovery. Such incidents rarely attract media attention when human administrators are responsible, which can skew public opinion in favor of human decision-making over AI.

Source

CISA says attackers are exploiting globally the two NetScaler RCE zero-days Citrix disclosed this weekend. The urgent alert expands the risk to eight vulnerabilities, including a third critical flaw that can bypass front-end security controls through HTTP request smuggling.

Source

Huawei is extending its open-source strategy from chips and servers to enterprise AI agents. At Huawei Connect 2026, the company presented an integrated stack combining the Atlas 960E SuperPoD, UnifiedBus networking, open-source software, and agent platforms designed to reduce dependence on proprietary AI ecosystems.

Source

Microsoft released KB5127216 and KB5125758 on September 22 for Windows 11 24H2 and 25H2, refreshing the components used during feature upgrades and system recovery. The releases extend Microsoft’s recent run of Windows 11 dynamic updates with new Setup security fixes and updated Windows Recovery Environment support.

Source

New Outlook will gain an optional spell check that interrupts sending when Microsoft detects possible spelling errors. The feature is scheduled to begin rolling out in October 2026, with wider availability expected by mid-December.

Source

Nvidia has released the free, open-source Open Agent Safety Platform, designed to contain autonomous AI agents before they can escape their assigned boundaries. The platform combines OpenShell sandboxing with Nvidia Sentry to enforce controls across agents, compute resources, and hardware—not just application prompts.

Source

Microsoft is expanding Microsoft Foundry with a model-agnostic agent platform designed to let organizations change AI models without rebuilding their applications. The update also adds native voice agents, resilient long-running tasks, on-demand tool discovery, and automated optimization based on production data.

Source