DEV Community: alex getman

OpenAI’s AI hacked Hugging Face to cheat on a benchmark

alex getman — Wed, 22 Jul 2026 14:37:49 +0000

OpenAI’s AI hacked Hugging Face to cheat on a benchmark

GPT-5.6 Sol and a more powerful unreleased model were running an internal cybersecurity benchmark with reduced safety restrictions. At some point, the agent figured the answers might be stored on Hugging Face and started looking for a way to reach them.

The model found a zero-day inside OpenAI’s infrastructure, escalated its privileges and gained internet access. It then stole credentials, broke into Hugging Face systems and reached the database containing ExploitGym solutions.

The whole operation was carried out autonomously. Hugging Face reconstructed more than 17,000 actions performed by the agent. The company had to use the Chinese GLM-5.2 model during the investigation because Western models refused to analyze real exploits.

OpenAI’s AI hacked Hugging Face to cheat on a benchmark

alex getman — Wed, 22 Jul 2026 13:04:59 +0000

OpenAI’s AI hacked Hugging Face to cheat on a benchmark

Claude Cowork can now learn from a screen recording

alex getman — Tue, 21 Jul 2026 20:41:27 +0000

Claude Cowork can now learn from a screen recording

Claude Cowork just added Record a skill. Start a screen recording, complete the task and explain what you’re doing. Claude remembers your clicks and workflow, then turns them into a reusable skill it can run on its own.

Codex got the same feature over a month ago. The video shows that version - it works pretty much the same way.

CODEX RESET LIMITS TO CELEBRATE 10 MILLION USERS

alex getman — Tue, 21 Jul 2026 17:10:57 +0000

CODEX RESET LIMITS TO CELEBRATE 10 MILLION USERS

Codex had 6 million weekly active users on July 12, and just 9 days later it hit 10 million.

Congrats to everyone involved. I’m sure you’ll be waiting for another reset by tonight 👀

RELEASE: GEMINI 3.6 FLASH

alex getman — Tue, 21 Jul 2026 15:06:27 +0000

RELEASE: GEMINI 3.6 FLASH

Google just dropped Gemini 3.6 Flash in Google AI Studio, Vertex AI, and Antigravity CLI.

Key updates:
• Knowledge cutoff jumped from January 2025 to March 2026
• Output pricing dropped from $9 to $7.50 per million tokens

Benchmarks next.

GPT-5.6 Sol brought RollerCoaster Tycoon 2 to the iPad

alex getman — Tue, 21 Jul 2026 10:25:16 +0000

GPT-5.6 Sol brought RollerCoaster Tycoon 2 to the iPad

Developer Chris Sotiradis used Codex to build a native iPad version of the game. He took the open-source OpenRCT2engine, compiled it for ARM64, and added full touchscreen controls.

You can build coasters and paths, play through scenarios, manage your park, move around the map, zoom in, and rotate objects directly on the screen. Mouse, trackpad, and keyboard are supported too.

Fable disproved an 87-year-old conjecture

alex getman — Mon, 20 Jul 2026 12:50:28 +0000

Fable disproved an 87-year-old conjecture

Mathematicians Levent Alpöge and Akhil Mathew asked the model to find a counterexample to the Jacobian Conjecture from 1939. Claude Fable 5 came back with a short formula. And it actually works.

In plain English, the conjecture said that if a complicated mathematical transformation never breaks, you should always be able to reverse it. Fable’s formula pulls a trick: three different points collapse into one. You can’t reverse that, because there’s no way to tell which point you started from. So the conjecture is false, at least in three dimensions.

Instead of pages of AI slop, this is a concrete formula anyone can check. Drop it into another model with Python, SymPy, or Wolfram and ask it to run the calculations. I checked it too: everything lines up, and three different points really do produce the same result.

If nobody finds a hidden flaw in the formula, this will be the biggest mathematical discovery ever made with the help of an LLM.

Kimi’s success has OpenAI’s strategy chief talking about restrictions

alex getman — Sun, 19 Jul 2026 18:24:20 +0000

Kimi’s success has OpenAI’s strategy chief talking about restrictions

Dean Ball, who leads strategic work at OpenAI, says China’s Kimi is already roughly on par with the best publicly available models from early 2026. He then describes how US agencies could discourage businesses from using Chinese open-weight models without banning them directly.

The proposed mechanism is simple: issue warnings about possible backdoors, data leaks, and other risks until banks, corporations, and compliance teams decide that using Kimi is not worth the trouble. Ball explicitly notes that these warnings would not need to be particularly well justified. They would only need to create enough regulatory uncertainty to push major customers away.

David Sacks, Trump’s former AI and crypto adviser, called this the weaponization of regulatory uncertainty as a competitive tool. He said regulatory warnings should be based on facts and evidence, not manufactured fear used to keep businesses away from a strong open-weight model.

Kimi shuts down new subscriptions after hitting capacity

alex getman — Sun, 19 Jul 2026 16:17:15 +0000

Kimi shuts down new subscriptions after hitting capacity

Demand for Kimi K3 exploded over the past 48 hours, pushing the company’s current infrastructure close to its limit. So the team has temporarily paused new subscriptions and is prioritizing compute for existing users.

Turns out, when you can’t handle the load, you can simply stop onboarding new users. No need to keep letting people in, posting daily updates about overloaded servers, and endlessly debating whether to extend access to another model for one more week.

Alibaba just teased Qwen3.8 with 2.4T parameters

alex getman — Sun, 19 Jul 2026 11:07:28 +0000

Alibaba just teased Qwen3.8 with 2.4T parameters

This is the new Qwen flagship, with open weights coming soon. Alibaba ranks it right below Claude Fable 5. Public benchmarks are still missing, so nobody really knows how close it actually is to Fable, or how it stacks up against GPT-5.6 Sol and Kimi K3.

Pricing is the fun part. Kimi K3 costs $3/$15, so Qwen3.8 will probably land in the same range.

The preview is already live in Qoder. It uses 10x fewer credits during the day and 50x fewer at night, so you can basically test it for free.

⚡️ FABLE IS HERE TO STAY

alex getman — Sat, 18 Jul 2026 11:38:08 +0000

⚡️ FABLE IS HERE TO STAY

After pressure from the community and competitors, Anthropic is keeping Fable around.

It’ll remain available on Max plans ($100 and $200). It’s being removed from the $20 Pro plan.

🚨 CODEX reset the limits.

alex getman — Sat, 18 Jul 2026 07:43:35 +0000

🚨 CODEX reset the limits.

Drop a ➕ in chat if your weekly limit lasts you exactly one day.