AI Daily Newsletter 16/07/2026: OpenAI Codex Encodes Agent Directives and Suno Caught Scraping YouTube Music

🔒 OpenAI Codex Encryption of Instructions Between Agents: When Programmers Are Left in the Dark

Recently, a significant change by OpenAI to Codex has sent shockwaves through the programming community. Since the beginning of June, the AI programming tool Codex has started automatically encrypting instructions that a primary agent transmits to its sub-agents. This means that developers have completely lost the ability to track the internal delegation process. With large models like GPT-5.6 Sol, this change creates a complete “black box” in multi-agent systems. Experts are concerned that the loss of ability to audit these intermediate commands will lead to unforeseeable security risks, as humans will no longer know what tasks the AI is assigning to each other.

OpenAI explains this action as a measure to secure the model’s privacy and optimize communication speed, preventing reverse engineering. However, the tech community feels that this is a step closer to a scenario where AI makes decisions without human supervision. If a sub-agent decides to delete a database due to a “misunderstanding” of the encrypted command from the primary agent, programmers will only be able to stand by and watch without being able to trace the error. This lack of transparency is raising a big question mark for the future of automation systems.

Source: The Decoder

🎵 Suno AI Exposed: Source Code Hack Reveals “Secret” Music Scraping from YouTube

Suno, one of the most prominent AI music startups today, has just faced a shocking data leak. A hacker used an employee’s login information to access the company’s internal source code. This hack has exposed a truth that Suno has always tried to avoid: they have been scraping decades of audio data from the YouTube platform to train their AI models. This is valuable evidence against Suno in the context of major record labels and artists suing them for large-scale copyright infringement.

Previously, Suno always argued that their use of data constituted “fair use” and refused to disclose details of their training dataset. Now, with code snippets proving automatic data collection from YouTube being made public, Suno’s legal stance is likely to be severely shaken. This incident once again sounds the alarm about the “gray area” of copyright in the AI creative industry. Can AI companies continue to freely use artists’ work to enrich themselves, or will the law tighten up? The answer will likely be decided soon in court after this scandal.

Source: TechCrunch

📱 Bonsai 27B: The Super Model of Reasoning Fits in Your Pocket

PrismML has made a significant breakthrough for on-device AI technology by launching Bonsai 27B. This is a complete open-source reasoning model with 27 billion parameters, but it has been ingeniously compressed to under 4 GB thanks to PrismML’s technology. With this extremely compact size, Bonsai 27B can run smoothly on a regular iPhone without needing an internet connection. According to PrismML’s internal tests, this super-compressed version of Bonsai retains up to 90% of the performance of the original uncompressed version.

This achievement opens up a new era where users can possess an intelligent AI assistant with deep logical reasoning capabilities right on their personal device without worrying about data leakage to the cloud or bandwidth costs. Running large models on phones used to be a significant challenge due to RAM and battery limitations, but PrismML’s advanced compression technique has solved that problem. Now, having a powerful “brain” of AI that can accompany you anywhere, even in areas without phone signal, has become a reality.

Source: The Decoder

⚔️ The Stick That Beats Its Own Back: OpenAI Uses AI to Attack Its Own Security System

In a latest report, OpenAI has revealed a unique but effective security method: using AI to self-attack its own system. The internal model named “GPT-Red” is trained through self-play to simulate network attacks. The result is truly astonishing as GPT-Red successfully found attack methods in 84% of the test scenarios, while human red teamers achieved only a 13% success rate. The vulnerabilities discovered by GPT-Red are immediately patched and used to strengthen the security barriers for commercial models.

This shows the speed and creativity of AI in finding loopholes far exceeds humans. However, it also opens up a worrying perspective: if a malicious entity possesses similar attack models like GPT-Red, traditional human defense systems will become useless. The arms race in AI security is no longer a battle between humans and machines, but a direct confrontation between intelligent algorithms.

Source: The Decoder

⌨️ OpenAI Sells Codex Mechanical Keyboard for $230 Amid Legal Storm

OpenAI has made a surprising move into the hardware market by launching a mechanical keyboard with LED lights under the Codex brand, priced at $230. This keyboard is designed to connect and optimize the AI agent application of the company. Notably, this launch occurs right in the middle of OpenAI being involved in a tense legal battle with Apple over allegations of stealing trade secrets related to hardware.

Many see this as a challenging step by OpenAI to assert that they are not intimidated by tech giants and are ready to build their own ecosystem of supporting hardware for AI. The Codex keyboard not only has a beautiful design but also integrates specialized shortcuts to help programmers interact faster with AI agents. Although the $230 price tag is considered high for a mechanical keyboard, for tech enthusiasts and AI developers, this is still a desirable tech gadget. This event also shows OpenAI’s ambition to not only dominate the AI software field but also to infiltrate every physical device on users’ workstations.

Source: TechCrunch