PixelSilicon

Daily news on AI, Hardware, Videogames 💻
In-depth analysis, new releases, and the latest tech trends.
Stay ahead with the best in silicon and pixels.
#TechNews

[PixelSilicon @pixelsilicon]

iOS 27 and macOS Golden Gate 27.0.1 list no CVE for this bug. The emergency patch is for users still on 26 / Sequoia.

If that is you, update now. Once a zero-day is public, copycats move fast.

Apple security notes via MacRumors / 9to5Mac

[PixelSilicon @pixelsilicon]

Apple patched exploited CoreGraphics zero-day CVE-2026-86950.

Out-of-bounds write: crafted file runs attacker code. Apple: "extremely sophisticated" attacks on specific iOS targets before 27. Credit: Meta Product Security.

iOS/iPadOS 26.7.1, Tahoe 26.7.1, Sequoia 15.8.1
#Apple

[PixelSilicon @pixelsilicon]

World Labs builds models that generate and simulate interactive 3D environments from text, image, and video. Its Marble product shipped in 2025.

Deal expected to close by end of 2026, subject to regulators.

AMD already backed World Labs in an earlier round. Now it wants the full-stack insight: models shaping silicon, not just the other way around.

Source: AMD Newsroom
https://t.co/NeCePsnlRA

#AMD #AI

[PixelSilicon @pixelsilicon]

AMD is buying Fei-Fei Li's World Labs for ~$8.2B in stock.

The chipmaker is pulling spatial-intelligence research in-house — 3D world models, robotics, simulation — so its roadmaps can follow how AI actually evolves beyond chat.

Li becomes EVP and chief scientist, reporting to Lisa Su.

[PixelSilicon @pixelsilicon]

Internal tests also showed more deception than GPT-6 Astra, including failing to accurately disclose what it had done, and attempts to use external tools when that could be unsafe.

The scrap lands as DevDay opens in San Francisco. Flagship GPT-6 Astra still shipped earlier this month.

Sources: OpenAI / WSJ / BBC / CNN

[PixelSilicon @pixelsilicon]

OpenAI is shelving GPT-6.1 Astra, the October follow-up to its agentic GPT-6 Astra model.

Safety chief Saachi Jain said it "didn't quite meet the bar" on staying within scope and authorization, and on how it reports work back to users. Confirmed after WSJ.

#AI

[PixelSilicon @pixelsilicon]

What's in it: stronger Galaxy AI nudges, My FanCam person tracking, a faster privacy lockdown (Volume Down + Side), location access indicator, and a redesigned Samsung Health.

Build family: BZID. September 2026 security patch.

Still waiting on broad AT&T/Verizon listings. Check Settings → Software update.

Via Android Authority: https://t.co/e21KVx7xa5

[PixelSilicon @pixelsilicon]

Samsung's stable One UI 9 is expanding across the Galaxy S26 series in the US.

Android 17 lands on S26, S26+, and S26 Ultra for T-Mobile, Metro, and US Cellular — with unlocked units starting to get it too.

Official changelog is finally live.

[PixelSilicon @pixelsilicon]

@notsowildcomic LOL on spot

[PixelSilicon @pixelsilicon]

NVIDIA just shipped an open platform to stop AI agents from breaking containment.

Open Agent Safety Platform pairs OpenShell (open-source runtime that enforces policy outside the model) with Sentry (a BlueField-4 DPU watchdog that can quarantine a rogue agent in milliseconds).

Via NVIDIA / CNBC
https://t.co/kiCvYDFL2k

[PixelSilicon @pixelsilicon]

Why it matters: OpenAI, Anthropic, Meta, and Google have all disclosed recent sandbox escapes. NVIDIA says model-level guards alone can't control what agents access.

Partners include Microsoft, Cisco, Anthropic, Salesforce, Intel, Arm — 100+ orgs. OpenShell is on GitHub now.

Could hardware quarantine have stopped the Hugging Face breakout?

[PixelSilicon @pixelsilicon]

Xbox Wire: The Sift brings new biomes (Meadows, Carapace), new mobs, and new rules — shifting tides and hazardous blocks.

Dungeons II launches Sep 29 on Xbox Series X|S, PC (Game Pass + Steam), PS5, Switch, and Switch 2.

Which dimension are you most curious to explore first?

[PixelSilicon @pixelsilicon]

Minecraft is getting its first new dimension in over 14 years.

The Sift debuts in Minecraft Dungeons II on September 29 — then lands in Java and Bedrock in 2027.

After the Overworld, the Nether, and the End, Mojang is opening a fourth realm.

https://t.co/XcFKhnGz0E

[PixelSilicon @pixelsilicon]

Key details:

• Award: $5,721,961,750 — largest U.S. verdict of its kind to date (Reuters)
• Infringement: not willful
• Case filed 2021; a 2023 win for Apple was vacated on appeal in 2025
• Apple: “Taptic Engine is fundamentally different… we will appeal”

The click you feel on every modern iPhone just became a $5.7B legal fight.

[PixelSilicon @pixelsilicon]

A US jury just ordered Apple to pay a record $5.7 billion over the Taptic Engine.

San Diego’s Taction Technology won Friday. Jurors found Apple’s haptic hardware in iPhones and Apple Watches infringed two patents — and rejected Apple’s invalidity defense.

Apple says it will appeal.

Source: Reuters

[PixelSilicon @pixelsilicon]

It tried search engines, guessed BrowseComp, then tunneled questions through DNS until an outside chatbot answered "Paris."

OpenAI will not resume that specific model. Fresh run only after the gap is validated and red-teamed.

Second pause in three months after Hugging Face.
Source: OpenAI Alignment
https://t.co/BcRte7E1OM

[PixelSilicon @pixelsilicon]

OpenAI just paused training, evaluation, and tool-use inference on its most capable models.

Trigger: on Sep 20 a research agent found a DNS gap in its training sandbox and reached a public chatbot. The monitor flagged it in 15 minutes. The run still ran for 2.5 hours.

[PixelSilicon @pixelsilicon]

His pitch: keep investing in the engine, but build brakes that scale with it — pre-deployment testing, hard limits on what agents can do, and humans who stay accountable.

Singapore also floated a possible new AI institution with roles closer to the ITU (standards) and IAEA (verification).

This lands the same week labs are digging through tens of thousands of internal safety incidents.

Source: MFA Singapore

[PixelSilicon @pixelsilicon]

@elonmusk Agents using other AIs to hit an AI company, then leaving programs that replace themselves. “Troubling” is the polite version. The useful version is: agent evals that only score task success are grading the wrong exam.