AI Weekly Briefing: 11-17 July 2026

AI security, consumer assistants and specialised hardware dominated the week. OpenAI deployed one model to attack another, Apple put its long-delayed Siri overhaul in public hands, and Thinking Machines Lab made its first open-weights release available for developers to customise.

OpenAI turns automated attacks into a training loop

OpenAI introduced GPT-Red, an automated red-teaming system that uses self-play to discover attacks against other AI models and agents. The company says the system can compromise nearly every model it tests and helped harden GPT-5.6 Sol against prompt injection. In one disclosed exercise, GPT-Red moved from simulated attacks to a production commerce agent, where it changed expensive products to low prices and cancelled other customers’ orders.

The work pushes adversarial testing towards a continuous engineering process rather than a periodic security review. That is increasingly necessary as agents gain permission to edit records, spend money and operate across business systems. GPT-Red also exposes the uncomfortable symmetry in model capability: stronger planning and tool use can improve both the defender testing an application and the attacker trying to subvert it.

Apple’s rebuilt Siri reaches the iOS 27 public beta

Apple released the iOS 27 public beta on 14 July, giving non-developers their first access to its redesigned AI-powered Siri ahead of the expected September launch. The assistant can use information from emails, photos and messages, respond to content displayed on screen, search through Spotlight and answer broader questions using world knowledge. Apple has also added a standalone Siri app while retaining voice, side-button and operating-system access.

The wider test is a significant reliability exercise across Apple’s installed base of roughly 2.5 billion active devices. Early developer builds handled tasks such as finding photos, summarising group messages and creating calendar entries from texts, but also produced errors and confused queries. The beta extends across Apple’s product range, although TechCrunch reports that Siri AI remains unavailable on iOS, iPadOS and watchOS in the European Union for now.

OpenAI ships Codex Micro as its wider hardware plans take shape

OpenAI released Codex Micro, a limited-run desktop control pad built with keyboard maker Work Louder. The $230 device has 13 mechanical switches, a joystick, dial and touch sensor. Six illuminated keys show the state of Codex tasks, while configurable controls can accept or reject changes, trigger push-to-talk, start workflows and adjust reasoning effort through the ChatGPT desktop app.

The launch is separate from OpenAI’s larger hardware programme with former Apple designer Jony Ive. Bloomberg reporting cited by The Verge describes a portable, screenless ChatGPT speaker with a camera, environmental sensors, smart-home controls and movable mechanical elements. The reported 2027 product would use OpenAI’s GPT-Live voice model and form part of a planned family of about five devices. Codex Micro is modest by comparison, but it provides an early test of whether persistent agent activity benefits from dedicated physical controls and status indicators.

Thinking Machines Lab debuts the open-weights Inkling model

Mira Murati’s Thinking Machines Lab released Inkling, its first model trained from scratch. The multimodal mixture-of-experts model has 975 billion total parameters, with 41 billion active for each request, and supports a context window of up to one million tokens. It was pretrained on 45 trillion tokens spanning text, images, audio and video. The company also previewed Inkling-Small, a lower-cost version with 12 billion active parameters.

Thinking Machines positions Inkling as a broad foundation for customisation instead of a benchmark leader. Its weights are available, reasoning effort can be adjusted to trade performance against cost and latency, and developers can fine-tune it through the company’s Tinker platform. The release gives the well-funded laboratory a concrete product and a clear initial market: teams that want a capable multimodal base model they can adapt to their own agentic, coding and domain-specific workloads.

Apple’s trade-secret lawsuit exposes an access-control failure

Apple’s lawsuit against OpenAI alleges that former system electrical engineer Chang Liu downloaded dozens of confidential hardware files after leaving Apple for OpenAI. According to the complaint, Liu used a previously unknown authentication flaw to retain access to Apple’s network storage and collected engineering presentations, technical specifications and unreleased product information over several weeks. Apple says it has fixed the vulnerability and found no evidence that other people exploited it.

The legal dispute will turn on evidence that remains contested, and OpenAI has said it is unaware of evidence supporting Apple’s complaint. The security failure described by Apple is more immediately concrete: a former employee allegedly retained working access to sensitive repositories after departure. For organisations connecting agents to internal systems, identity deprovisioning, short-lived credentials and auditable permissions become more important as automated tools make it easier to locate and extract large collections of data.

Cars24 reports measurable returns from voice and chat agents

Indian automotive marketplace Cars24 says its OpenAI-powered voice and chat agents now handle more than one million conversation minutes each month. The company reports that the systems recover 12 per cent of leads that would otherwise have been lost, while teams are extending agentic workflows beyond customer conversations into internal operations.

The figures offer a useful production-scale reference for enterprise agent deployments. Conversation volume is easy to report, but recovered leads connect the system to commercial value. Cars24’s case also reflects a broader shift from isolated support chatbots towards agents embedded across sales and operational processes, where integration quality and controlled access often determine the result more than model benchmarks.

Also worth noting

London-based Applied Computing raised €17.4 million to expand its foundation AI systems for energy operators and open a US office. The round shows continued investor interest in industry-specific models that can work with operational data and constraints rather than serving as general-purpose assistants.

OpenAI also published details of its approach to age-appropriate ChatGPT access for teenagers, including parental controls, learning tools and expert partnerships. The announcement arrives as providers face growing pressure to demonstrate that consumer AI safeguards account for age and usage context rather than applying one policy to every user.