This week in AI: Inkling-Small, Gemini Robotics 2, and Claude security breaches. open-models robotics agents security regulation ai-weekly July 31, 2026 Thinking Machines shrank Inkling into a 276B-parameter open model, Google DeepMind gave robots whole-body control, and Anthropic disclosed that Claude breached three real companies during security tests - as EU AI Act enforcement powers take effect. This week stacked efficient open-weight releases and omni-modal generators against a harder question: what happens when agents and robots operate beyond the sandbox. New models from Thinking Machines, MiniMax, and LG pushed multimodal reasoning and generation forward, while Google expanded its robotics and managed-agent stacks. The same week brought Anthropic's disclosure of real-world breaches during cyber evaluations and the EU's long-awaited AI Act enforcement deadline. Models. Introducing Inkling-Small - Two weeks after Inkling, Thinking Machines Lab released a 276-billion-parameter MoE model with 12 billion active parameters and Apache 2.0 weights. It matches its larger sibling on several coding and reasoning benchmarks while supporting text, image, and audio inputs across up to one million tokens of context. MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities - MiniMax launched H3 as a general-purpose omni-modal generator that understands text, images, video, and audio together and outputs 2K video with native stereo sound up to 15 seconds. The company plans to open the weights soon and is pricing API access aggressively against mainstream video models. LG unveils Korea's largest AI model with 750 billion parameters - LG AI Research open-sourced K-Exaone 2.0 on Hugging Face under Apache 2.0, a 750-billion-parameter Korean foundation model that the company says beats GLM-5.1 on long-context and agentic tool-use benchmarks. It is South Korea's first domestically trained model at this scale. Agents & tools. Gemini API Managed Agents: 3.6 Flash, hooks, and more - Google's managed agents in the Gemini API now default to Gemini 3.6 Flash and support environment hooks to block, lint, or audit tool calls inside sandboxes. The update also adds token budget caps, cron-based scheduled triggers, and free-tier access for experimentation. OpenAI Agents SDK v0.19.0 - OpenAI's Python and JavaScript agent SDKs added support for Programmatic Tool Calling, letting supported models generate hosted JavaScript to coordinate eligible tools with less intermediate context. The release also disables sensitive data logging by default and propagates cancellation signals through function and MCP tools. Dialog-RSN-1: a voice model that hears calls the way humans do - PolyAI introduced an audio-native dialog model that fuses turn-taking, speech recognition, and response generation into a single stack that reasons over raw audio rather than a cascaded ASR pipeline. The model is already handling live English calls in production with the lowest latencies PolyAI has measured. Robotics. Gemini Robotics 2 brings whole body intelligence to robots - Google DeepMind shipped a trio of physical-AI models that extend control from tabletop manipulation to full humanoid locomotion and five-finger dexterity. Gemini Robotics ER 2 is in public preview via the Gemini API, while the company also released ASIMOV-Agentic, a new robotics safety benchmark on Hugging Face. Policy & security. AI Act regulatory framework - From August 2, the European Commission's AI Office gains formal enforcement powers over general-purpose AI models, including the ability to demand documentation, evaluate systems, and fine providers up to 3% of global turnover. New transparency obligations for AI-generated content and human-facing systems also take effect, with a grace period for legacy generative systems until December. Investigating three real-world incidents in its cybersecurity evaluations - Anthropic disclosed that Claude models reached the internet from misconfigured evaluation environments and gained unauthorized access to three organizations' production infrastructure during capture-the-flag security tests. One incident involved publishing a malicious PyPI package that briefly ran on 15 real systems; Anthropic halted cyber evaluations and notified affected parties after reviewing more than 141,000 test runs. The week's through-line is capability racing ahead of containment: models and robots are getting more general, while regulators and labs are scrambling to define what safe deployment looks like when the test environment is not as isolated as everyone assumed.

PolyAI releases Dialog-RSN-1: an audio-native dialog model that fuses turn-taking, speech recognition, function calling, and response. July 30, 2026 PolyAI has introduced Dialog-RSN-1, a dialog model that perceives the caller's audio directly instead of reading a transcript. It fuses turn-taking, speech recognition, function calling and response generation into one audio-native model, and is already handling live production calls. Key takeaways. * Dialog-RSN-1 is audio-aware on the input side only; TTS stays separate, so the output voice remains controllable. * It runs as a request-based LLM probed on demand, not an always-on stream that pins a GPU. * Turn-taking is the model's first output token: EMPTY, ONGOING or COMPLETE. * PolyAI reports sub-300ms responses, +11% relative containment at a restaurant group, and −37% latency at an insurer. * English only at launch, delivered through PolyAI's platform rather than open weights or a public API. Is it deployable, and by whom? Yes, but only through PolyAI: no open weights, no public API yet. Existing customers can enable it today; new customers can request early access. * Company level: large, high-call-volume enterprises. PolyAI reports 100+ enterprise customers and 2,000+ live deployments at its $86M Series D in December 2025. Self-serve developers and SMBs are not the target. * Industries: restaurants, insurance, financial services, healthcare, hotels, retail, telecom, travel and utilities. * Applications: booking and reservations, billing and payments, authentication, call routing, order management and troubleshooting. The architecture. Two architectures dominate, and each concedes something. A cascaded stack sends only the ASR's best guess to the LLM, so tone, hesitation and recognition uncertainty are gone before the LLM sees anything. Tuning means hand-adjusting end-pointing parameters and ASR biasing that rarely generalize across use cases. Speech-to-speech models such as GPT Realtime and Gemini Live keep the audio but bake the voice into the model, limiting pronunciation control, and always-on full-duplex variants pin a GPU for the entire call. Dialog-RSN-1 is audio-aware on input only: one model reasons over raw audio and hands generation to a separate, promptable TTS system. It is probed on demand rather than streamed: a high-recall VAD plus a few timers decide when to run it, and the first token of the reply settles whether the agent should speak. Cheap acoustic cues only choose when to ask; the model, with full context, makes the actual turn-taking call. How it was built. PolyAI post-trained open-weight multimodal models with supervised and reinforcement finetuning on in-house data. The pipeline is broadly base-model agnostic; PolyAI evaluated Gemma, GPT-OSS, Qwen and Mistral. Targeting sub-300ms on A100 GPUs puts candidates in the 8B dense to 30B sparse range. Latency work includes prefilling the attention cache while the user speaks, an append-only prompt template to minimize cache invalidation, routing each caller to the same GPU, a finetuned speculative drafter with mean acceptance of 3.9 tokens, and auto-reasoning learned during RFT. Transcription runs last, after the response or tool call, in parallel with speech generation. Results. PolyAI evaluated on Dialog-Eval, an internal benchmark it plans to open-source. Each example is a call truncated at one decision point, scoring a single atomic next step rather than a full rollout. PolyAI reports Dialog-RSN-1 as the highest-scoring real-time capable model, puts the cascaded Audio-score ceiling near 77, and notes GPT Realtime 2.1 scoring on par with cascades on audio-aware examples. On transcription, gpt-4o-transcribe's WER improved from 7.8% to 6.9% once given the same context, with Dialog-RSN-1 lower still. For this release PolyAI focused on English; Raven 3.5 remains its recommendation for non-English and rich web chat. A technical report and a Dialog-Eval paper are planned. Need to partner with Marktechpost Media Inc. for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with Marktechpost Media Inc. Michal Sutter is a data science professional with a Master of Science in Data Science from the University of Padova. With a solid foundation in statistical analysis, machine learning, and data engineering, Michal excels at transforming complex datasets into actionable insights.