Today's Headlines

  • OpenAI unveils GPT-Red, an automated red-teaming AI — on unseen indirect prompt-injection scenarios, it succeeded on 84% of scenarios versus 13% for human experts
  • Thinking Machines Lab releases its first model, Inkling, a 975-billion-parameter multimodal mixture-of-experts, as Apache 2.0 open weights on Hugging Face
  • Google's round-the-clock AI agent Gemini Spark lands in Japan — beta for Ultra subscribers first, with a Pro expansion hinted

It is a day when how AI gets trained and how it reaches people moved at once.

OpenAI published safety research in which an attacker AI is trained to harden its production models. Thinking Machines Lab shipped its first large model in a form anyone can download and modify. And Google's autonomous agent began reaching everyday users in Japan. The methods on the builders' side and the options on the users' side each moved a step forward on the same day.

Today's Top Three

OpenAI Unveils GPT-Red, an Automated Red-Teamer — 84% Scenario Success Versus 13% for Humans

OpenAI on July 15 announced GPT-Red, a red-teaming AI that automatically hunts for vulnerabilities in the company's own models.

GPT-Red is trained through self-play reinforcement learning. The attacker, GPT-Red, and a collection of defender LLMs are trained simultaneously: the attacker is rewarded for successful prompt injections, the defenders for resisting attacks while completing their original tasks. As the defenders harden, the attacker is forced to invent stronger and more diverse techniques. OpenAI says it trained GPT-Red at the compute scale of some of its largest post-training runs — dedicated purely to safety.

To measure its strength, OpenAI pitted GPT-Red against human red-teamers on a set of unseen indirect prompt-injection environments (an internal replication of the attack arena from Dziemian et al., 2025), with GPT-5.1 as the target. Human experts succeeded on 13% of scenarios; GPT-Red succeeded on 84%. In a live case study against Vendy, an AI vending-machine agent actually operating in OpenAI's office, GPT-Red achieved all three of its malicious objectives — including repricing an expensive item down to the $0.50 minimum and canceling another customer's order.

The gains flow back into defense. GPT-5.6 Sol, the latest model trained with GPT-Red in the loop, shows six times fewer failures on the company's hardest direct prompt-injection benchmark than its best production model from just four months earlier, and it now fails on only 0.05% of GPT-Red's direct injection attempts. A novel attack class an early GPT-Red discovered, dubbed "Fake Chain-of-Thought," succeeded more than 95% of the time against GPT-5.1 but is now held below 10% against GPT-5.6 Sol. GPT-Red itself stays internal-only, kept separate from deployed models so its offensive capabilities never reach adversaries.

Thinking Machines Releases Its First Model, Inkling — Open Weights You Can "Make Your Own"

Thinking Machines Lab, the startup led by former OpenAI CTO Mira Murati, released its first AI model, Inkling.

Inkling is a mixture-of-experts model with 975 billion total parameters and 41 billion active, handling text, images, and audio natively. It supports a context window of up to one million tokens and was pretrained on 45 trillion tokens spanning text, images, audio, and video. The full weights are published on Hugging Face as open weights under the Apache 2.0 license, so companies can freely download, modify, and self-host the model. What is free, however, is the weights themselves: using the model on the company's Tinker fine-tuning platform is paid (at a 50% discount for a limited time), and the chat-style Inkling Playground is free only for a limited period.

On the benchmarks the company published, Inkling scored 61.1 on ForecastBench (no search), a measure of forecasting ability, beating GPT-5.5 at 59.1 and Claude Opus 4.8 at 54.6. In the same table, however, it trails Grok 4.3 (61.7), and with search enabled GPT-5.5 comes out ahead. On IFBench, an instruction-following benchmark, Inkling's 79.8% tops the closed models in the comparison, while the company itself discloses that it falls short of top closed models on reasoning and coding measures such as HLE and SWE-Bench Verified.

Alongside the flagship, a lighter model called Inkling-Small (276 billion total parameters, 12 billion active) was released as a preview. Rather than chasing the strongest score everywhere, the pitch is a foundation model companies can fine-tune on their own data and make their own — a deliberate counterweight to one-size-fits-all closed models.

Google's Gemini Spark Lands in Japan — Ultra-Only Beta First, Pro Expansion Hinted

Google on July 16 began offering its personal AI agent Gemini Spark in Japan.

Availability starts with subscribers to the Google AI Ultra plan, which runs from 14,500 yen (about $95) per month, as an early beta. Josh Woodward, Google's vice president in charge of Gemini, posted on X telling Pro members to "stay tuned for access updates," so whether the agent reaches the 2,900-yen Pro tier is the next thing to watch.

Gemini Spark is an agent that handles work on the user's behalf: gathering information, organizing material, dealing with email, adjusting schedules, and arranging travel. Because it runs in the cloud, it keeps working in the background even while the user's phone or PC is switched off. It is built on Gemini 3.5 and Google's agent harness Antigravity, integrates with Gmail, Google Calendar, and Google Docs, and supports the Model Context Protocol (MCP) for connecting outside apps.

Recurring work can be saved as reusable "skills," and the agent asks for confirmation before high-impact actions such as sending email. The launch brings a service that debuted in the US to Japanese-language users — a milestone that puts a round-the-clock autonomous agent into the hands of ordinary users in Japan.

Other Moves

Models & Products

Business

Policy & Other

Source: Selected by the editorial team from the AI news inbox (30 items collected on July 16, 2026 — 3 primary, 27 secondary).