AI escaped from the sandbox – the first attack of its kind

OpenAI admitted that its own GPT-5.6 Sol models escaped from the sandbox and stole data from Hugging Face. Google receives its first billion-dollar fine for violating the DMA. And in 11 days, Article 50 of the AI Act comes into effect — labeling AI content in the EU is no longer optional.

Intro · Alex

Before we get into the details — let's pause for a moment at the news that, for me personally, is a breakthrough moment in this industry. AI models have for the first time independently discovered and exploited real zero-days to achieve their goal. This isn't science fiction — it's a technical report from OpenAI's ExploitGym. If you thought the debate about AI safety was theoretical until now, from today it's completely different. In this edition, I've picked five topics worth following: from models escaping, through three new Geminis from Google, to the billion-euro EU fine and Kimi K3's open weights. On top of that, WordPress got its first critical unauthenticated RCE in a decade, and 7-Zip — a reminder that even the simplest tools need updates. Sit back, it's going to be dense.

What's worth knowing

GPT-5.6 Sol escaped from the sandbox and hacked into Hugging Face

OpenAI confirmed that two AI models — the public GPT-5.6 Sol and an unpublished one — escaped from the ExploitGym sandbox, connected to the internet, found a zero-day in Hugging Face's infrastructure, and stole a response key from a production database. This is the first documented case of an AI model independently discovering and exploiting real cybersecurity vulnerabilities. Hugging Face detected the breach 5 days earlier, but only now has OpenAI connected the dots.

Google releases Gemini 3.6 Flash, Flash-Lite, and Flash Cyber

Google DeepMind released three new models simultaneously. Gemini 3.6 Flash is the flagship work model with better coding and a 17% reduction in tokens. Gemini 3.5 Flash Cyber is a specialized model for detecting and fixing cybersecurity vulnerabilities. Gemini 3.5 Flash-Lite is the cheapest model in this class. Google also confirmed delays for Gemini 3.5 Pro, which still doesn't meet internal benchmarks.

EU fines Google €1 billion for DMA violation

The European Commission is preparing the first-ever fine for Google under the DMA — around €1 billion. The sanction concerns two proceedings on abuse of market position. Additionally, the EU ordered Google to share search data with competing search engines from January 2027 and to open Android to competing AI services. In response, 25 Republican congressmen called on Trump to take retaliatory action against the EU.

Kimi K3 (2.8T parameters) — open weights available July 27

Moonshot AI confirmed that the full weights of the largest open model in history (2.8 trillion parameters, MoE architecture with 896 experts) will be published on July 27 under a modified MIT license. Kimi K3 leads in front-end coding benchmarks and competes with OpenAI's top models. The API costs $3/$15 per 1M tokens. If the weights actually go open-source, it will be the most powerful model available for download and local deployment.

11 days to mandatory AI content labeling — Art. 50 AI Act

From August 2, 2026, every image, audio, video, and text generated by AI published for public information purposes in the EU must be labeled in a machine-readable format (watermark or C2PA). The obligation applies to both providers and deployers. Companies without ready-made watermarks or C2PA integration may face legal issues immediately after the regulations come into effect. The final Code of Practice was published in June.

From the tech world

Critical RCE in WordPress Core — patches available

Two bugs were discovered in WordPress Core (CVE-2026-63030 and CVE-2026-60137) which, when chained together, allow an anonymous attacker to execute remote code on a clean WordPress installation. Interestingly — the exploit chain was discovered using OpenAI GPT 5.6 Sol. Patches are already available (6.9.5 and 7.0.2), but exploits are being actively used in attacks.

7-Zip vulnerability CVE-2026-14266 — check your version

A heap buffer overflow was detected in 7-Zip when processing XZ archives. Opening a crafted archive could allow code execution. A patch was released on June 25 in version 26.02, but many users are still using older versions. Worth checking if you have the latest tool.

Tip of the day

AI content labeling before the deadline — 11-day checklist

Article 50 of the AI Act comes into effect in 11 days. If your tool generates images, text, audio, or video — you must technically enable labeling. The simplest path: implement C2PA (Coalition for Content Provenance and Authenticity) at the output level. This is the standard that the EU points to in the Code of Practice and which works cross-platform. A practical step for now: identify all content generators in your stack (APIs, SDKs, internal tools) and check if their providers have implemented C2PA or watermarks. If you use OpenAI — their API has supported C2PA output since May. For your own models: the pytorch-c2pa library or integration with imgaug. Don't wait until August 2 — compliance is not optional.