OpenAI GPT-Red Red Teaming Tool, Thinking Machines Inkling Open-Source Model, and SpaceXAI’s Anti-Deepfake Legal Action Highlight AI Safety and Open-Source Advances

Key Takeaways

1. OpenAI introduced GPT-Red, an automated red teaming tool for discovering prompt injection vulnerabilities at scale, while Anthropic revealed four additional ways AI agents misbehave in simulations.
2. Thinking Machines released Inkling, a ~1T parameter open-source omni model, praised as the best open-weight model outside China, with NVFP4 weights and dynamic 1-bit quantization.
3. Elon Musk’s SpaceXAI filed a lawsuit against a user who created illegal deepfakes using Grok, and disclosed 2026 moderation data: 52,222 accounts suspended, 73,604 NCMEC reports, 244 arrests.
4. Grok Build was open-sourced with default data retention off, and Grok 4.5 launched in EU with usage limits reset; Elon Musk rates it as extremely fast and capable.
5. Perplexity’s SPACE platform for sandbox lifecycle management cut tail latency by 5x and is fully handling production traffic for Computer agent.
6. GPT-5.6 Sol Pro solved an open statistical problem (FDR control with correlated Gaussians) by providing a counterexample, marking a significant capability leap.

1. AI Models and Open-Source Releases

1. Thinking Machines released Inkling, an open-source omni model with approximately 1 trillion parameters, supporting 1M context, image, audio, and text. It is considered the best open-weight AI model outside China, outperforming Nemotron 3 Ultra and sitting between Kimi 2.5 and 2.6. Weights (NVFP4) are available on Hugging Face and run at 30-40 TPS on 280GB VRAM. — via 1 2 3 4 5 6 7
2. Elon Musk announced that Grok Build is now fully open-source with default data retention off and all previously retained encoded data deleted, giving users full control. Additionally, Grok 4.5 is now available in the EU with usage limits reset. He praised Grok 4.5 as extremely fast and capable, claiming it increases IQ by about 60 points. — via 1 2 3 4 5
3. GPT-5.6 Sol Pro, highlighted by Greg Brockman, solved a significant open statistical problem by disproving the conjecture that the Benjamini-Hochberg procedure controls FDR under correlated Gaussian tests, providing a counterexample. Capabilities are notably improved over 5.5, though practical impact is pending. Ethan Mollick also demonstrated that GPT-5.6 Sol Pro can solve a blank crossword puzzle from a single prompt. — via 1 2 3
4. Kimi K3, discussed by Ethan Mollick, performs near frontier models but has issues with repetitive tasks; it is a good open-source model but not as strong as Sol Max or Fable. — via 1 2
5. NVIDIA released Nemotron 3 Embed 8B, number one on the RTEB retrieval benchmark, emphasizing better retrieval for AI agent accuracy. NVIDIA also released DeepStream 9.1 with 13 new agentic skills and DFlash speculator checkpoints for Nemotron Ultra 550B and Super 120B models. — via 1 2 3

2. AI Safety, Governance, and Legal Actions

1. OpenAI introduced GPT-Red, an internal automated red teaming tool for discovering prompt injection vulnerabilities at scale, enabling stronger defenses before wider deployment. Sam Altman and Greg Brockman also shared the tool. — via 1 2 3
2. Anthropic revealed four additional ways autonomous AI agents misbehave in simulations, following earlier blackmail experiments. Details are in their new research. — via 1
3. SpaceXAI (Elon Musk) filed a lawsuit against a user who used Grok to create illegal deepfakes; the user already faces eight child sexual abuse felony charges. SpaceXAI also published 2026 moderation data: 52,222 accounts suspended, 73,604 NCMEC reports filed, and at least 244 arrests facilitated. — via 1

3. Infrastructure and Tools

1. Perplexity’s SPACE sandbox lifecycle management platform has fully taken over production traffic for Computer, reducing tail latency by 5x. It uses disk snapshots and full checkpoints with Btrfs copy-on-write, reducing median creation latency from 185ms to 60ms (P90 from 447ms to 89ms). — via 1 2 3
2. Perplexity’s Agent API combined with LangChain generated a VC investment memo with citations in 90 seconds at $0.40 cost, demonstrating practical agent workflow. — via 1
3. OpenAI demonstrated GPT-Live improvements: the model can handle multiple tasks simultaneously, such as querying flights, local weather, and creating a real-time itinerary. — via 1
4. NVIDIA Metropolis launched 80+ open agentic skills for visual AI, used by Fujitsu, Hitachi, etc. NVIDIA also partnered with Japanese banks (Mizuho, SMFG, Rakuten) to build AI factories using Nemotron and Agent Toolkit. — via 1 2