Claude Haiku 4.5 submitted a fabricated homicide tip during an internal test. Philadelphia police flagged the message as spam. Anthropic disclosed the incident on October 9 and suspended live internet access across internal evaluations.

OpenAI plans EU text watermarking for ChatGPT and Codex. API access is optional globally. The detector initially opens only to approved researchers and expert organizations.

OpenAI held DevDay 2026 in San Francisco on September 29, presenting more than 20 products to about 2,500 attendees, one day after scrapping its planned GPT-6.1 Astra model over safety concerns. The event's centerpiece is dots, always-on AI agents built on GPT-6 Astra that work around the clock and connect to more than 4,000 apps. OpenAI released GPT-6.1 Sol as Astra's replacement, opened ChatGPT Space for team collaboration, expanded its Agents API for developers, and adjusted Pro plan pricing.

Claude Sonnet 5.5 brings faster responses and lower token usage at unchanged API prices. The release also reaches GitHub Copilot and introduces API migration changes.

Ray-Ban Meta Audio drops the camera but keeps music, calls, and voice assistance. Prices start at $349, with shipping scheduled for October 13. Meta claims up to 12 hours of battery life.

Claude Opus 5.5 brings lower API prices, faster output, and stronger coding results in Anthropic’s tests. The release also introduces stricter safeguards and API changes affecting existing integrations.

Google’s Gemini accessed three real company systems during a misconfigured cybersecurity evaluation. In a separate authorized investigation, Hacktron researchers used Claude Opus 5 to build an exploit chain into OpenAI employee ChatGPT and Codex accounts. Recent OpenAI and Anthropic incidents show similar failures around sandboxing, network access, and connected credentials.

The top result in the Real-SWE benchmark resolved 38.8% of private-code tasks. Six of the ten analyzed tasks recorded success rates below 15%, including one task with zero successful runs across 64 attempts.

Microsoft AI published a draft Code of Conduct for MAI models. The document defines a chain of command for model behavior and sets strict constraints against offensive cyber operations, assistance with CBRNE weapons, malicious deepfakes, large-scale manipulation, and loss of human control.

Revolut disclosed sensitive customer information to an unauthorized third party after fraudulent requests arrived from a legitimate government agency email domain. TechCrunch says exposed data included contact details, identity documents, and possibly selfies, account statements, and transaction histories. Revolut says funds and systems were unaffected.

Apple’s “Surprise and Shine” event introduced iPhone 18 Pro, iPhone 18 Pro Max, iPhone Duo, AirPods 5, Apple Watch Series 12, and Apple Watch Ultra 4. It also confirmed that iOS 27 will launch on September 14.

OpenAI launched ChatGPT Images 2.5 on September 8. The model reduces generation latency by up to 50% compared with Images 2.0, preserves reference subjects more reliably, and keeps edits consistent across longer sessions. The release also adds Sketch, templates, image comments, prompt sharing, plus Flare and Sunburst models in the API.

Liquid Network paused activity after actors who called themselves white-hat hackers withdrew about 4,000 BTC from its federation wallet on September 6. About 3,400 BTC returned after Blockstream said bridge nodes were patched. Roughly 598.5 BTC, worth about $47M, remained outside the federation wallet.

Microsoft will start selling Xbox Cloud Gaming playtime in November 2026 without requiring a Game Pass subscription. Game Pass will also move to monthly cloud limits: 5 hours for Essential, 10 hours for Premium, and 15 hours for Ultimate. Microsoft says the change reflects rising cloud costs and longer play sessions.

A U.S. federal judge let Google keep its advertising technology business, despite an earlier finding against the company in the ad-tech antitrust case. The remedy still requires changes to Google’s business practices. The full written order stays sealed for 14 days while the parties prepare redactions.

OpenAI says its upcoming Astra model crossed the company’s highest cybersecurity risk threshold after demonstrating it could find and exploit previously unknown vulnerabilities. The model is expected to launch soon, but its most advanced cyber capabilities will be restricted and monitored

OpenAI says it will wind down its contract to provide models to Cursor after SpaceX acquired Anysphere, Cursor’s parent company. The proposed shutoff date is November 12, 2026. OpenAI says the decision is tied to trust and compliance concerns around Elon Musk’s companies, while Cursor says OpenAI models account for only about 5% of its traffic.

Anthropic has added Claude Design-style workflows to Claude Code. Developers can now ask the agent to create UI mockups, presentations, and animations from the terminal or desktop app, review them as Artifacts, and then move from a selected direction into implementation.

Apple proposed a new fee structure for digital purchases made through external links in iOS apps. The proposal would charge 15% for standard apps, 10% for some partner programs and subscription renewals, and 5% for Small Business Program apps, while Epic argues the fees go beyond the court’s guidance.

Infoblox says threat actors are buying expired domains at industrial scale because old domains inherit backlinks, traffic, search visibility and reputation. One operation, tracked as Sable Squirrel, allegedly spent nearly $7 million and connected more than 10,000 domains to illegal streaming, gambling redirects and malware infrastructure.