Cisco released AI Skill Scanner, a security tool for AI agent skills that detects prompt injection, data exfiltration, and malicious code patterns. It combines multiple engines such as static analysis, LLM evaluation, behavioral dataflow analysis, and provides CI/CD integration and pre-commit hooks.

Nashra is an all-in-one platform combining email newsletters, blogging, and landing pages, with AI-powered features to help creators convert followers into clients. Launched today on Product Hunt.

The article argues that AI agents are becoming the primary interface for products, surpassing traditional UIs. It discusses the evolution from MVPs to 'slop', the impact on data platforms like Snowflake and Databricks, and presents Hopsworks' strategy of staying as the data layer for agents.

OpenAI is scrambling to fix a flawed logging implementation in its Codex coding agent that has been shortening the lives of users' SSDs with excessive data writes, potentially costing millions of dollars in lost drive endurance.

OpenAI and Broadcom introduce Jalapeño, a custom AI chip built for LLM inference to improve performance, efficiency, and scale across AI systems.

In the June 23, 2026 daily AI report, agents are moving beyond chat interfaces into physical control. Anthropic taught a robot to walk 37x faster with Claude Opus 4.7; Nvidia launched a spatial reasoning framework; Tesla advanced modular data centers; Sakana AI introduced Fugu for multi-model orchestration. The control and safety layer is racing to catch up, while 'cognitive debt' warns of human skill atrophy.

Ruby is an AI tool that listens in on calls and suggests the right questions in real time, helping you capture key information. Ideal for user interviews, sales calls, and investor meetings.

VoltanaLLM introduces a feedback-driven frequency controller and state-space router for prefill/decode disaggregated LLM serving, achieving up to 36.3% energy savings while maintaining latency SLOs.

Swimio combines AI coaching, advanced Apple Watch swim tracking, and personalized workout generation for swimmers. Track pace, SWOLF, stroke efficiency, and heart rate, and get insights after every session.

Japanese AI startup Sakana has posted benchmark results and beta tester impressions saying that its Fugu models compare favorably with or beat the likes of Fable 5, Gemini 3.1 Pro, and GPT-5.5. The company says Fugu carefully chooses when to use other frontier AI models like Claude and Gemini on specific tasks, although it doesn’t tell users which model was used for what.

Fork.ai is a branching AI research workspace that lets you ask a question and receive an answer split into sections. You can dive deeper, highlight, and branch from each section, turning detours into nodes on a mind map.

Garment workers in India were asked to wear head-mounted cameras at work without explanation, sparking concerns about AI replacing jobs.

darktable 5.6 ships its first AI features: an AI object mask that turns a single click into a vector mask around any subject, and a neural restore module for ML-based denoise and upscale. All models run locally on CPU or GPU with no cloud dependency. The article provides hands-on walkthroughs, tips, and honest caveats.

A community-created open-source library of 817 production-grade cybersecurity skills across 29 domains, mapped to six industry frameworks (MITRE ATT&CK v19.1, NIST CSF 2.0, etc.). AI agents can use these skills to perform threat detection, digital forensics, and more, with senior analyst-level decision-making capabilities.

NavWM is a unified navigation world model that integrates latent world reasoning, multimodal action prediction, and controllable visual generation. By introducing an anchor-based multimodal trajectory forecasting framework, it generates a diverse action space and uses visual foresight for robust closed-loop planning. Experiments show significant improvements in high-fidelity future state generation and zero-shot navigation success.

DynaWM improves bipedal-wheeled robot locomotion on continuous stairs by using a world model regularizer for terrain encoding and a momentum target encoder for stable distillation, enabling smoother and more adaptable movement in simulation and real hardware.

MinInter selects source demonstrations requiring the least interpolation to generate higher-quality synthetic data for imitation learning. Experiments on 12 manipulation tasks show consistent improvements in data generation and policy success rates, with largest gains on contact-rich, long-horizon tasks.

SPACE framework uses Cartesian state delta as a universal action representation, with State Prediction and Adaptive Command Execution to address issues in behavior cloning across robots with different dynamics. Experiments show it outperforms direct command prediction and remains robust under dynamics shifts.

TurboMPC is a differentiable MPC solver that runs entirely on the GPU, supporting state and control inequality constraints, implicit integrators, cross-time-coupled costs, and slack variables. It achieves up to 15× and 58× speedups over state-of-the-art CPU and GPU differentiable solvers, respectively, and scales to planning horizons over 8000 knot points. Deployed on a full-scale car for minimum-time racing, GPU-accelerated Bayesian optimization tuning yields significantly faster driving.

This note describes an integration of the sim-to-real performance estimate with betting (from Chen et al.) and the safe anytime-valid inference (from Ramdas et al.), using scaled simulators to produce efficient, reliable certificates for mean estimates, especially valuable in robot performance testing.

This paper introduces TOLD, a real-time edge-level adaptation framework that updates interaction topology weights online to minimize formation distortion, outperforming conventional node-level robust controllers. Theoretical analysis, simulations, and hardware experiments on Crazyflie 2.0 quadrotors demonstrate significant distortion reduction (over 62% for OGF).

Reinforcement learning can train bimanual dexterous hands to play piano in physics simulation with high note accuracy, but for high-DoF hands, relying solely on task rewards or IK inversion often leads to unnatural postures and joint overextension. The proposed Adversarial Posture Regularization (APR) uses a small amount of casual human playing data to match the posture distribution of the policy with a human prior via an adversarial objective, encouraging more human-like hand shapes. The authors collect and release unstructured hand motion data using a consumer-grade Meta Quest 3 and retarget it to the Shadow Hand. APR achieves significantly better performance than prior methods on human-likeness metrics (cPSI, BSE, FAC) and visual quality.

This workshop report captures discussions from the Lorentz Center Workshop "Engineering Reliable Autonomous Systems" (ERAS), held June 10–14, 2024. It focuses on verification and validation, real-world engineering, and safe software architectures for autonomous systems, resulting in a catalogue of challenges and a roadmap to solutions. Some challenges can be addressed by existing academic techniques not yet widely adopted in practice; others require further research.

This paper introduces FEARL, a framework that decomposes robot policies into a large controller and a small safety module, enabling formal verification of safety-critical properties while preserving the expressive power of foundation models. Experiments in simulation and on a physical robot demonstrate its effectiveness.

Advances in generative AI have made image falsification highly realistic, demanding trustworthy authentication systems. Existing forensic detectors lack interpretability, while vision-language models (VLMs) provide explanations but cannot exploit forensic traces for reliable detection. We propose Forensic Knowledge Graphs (FKGs), a unified framework integrating forensic evidence extraction, structured reasoning, and human-interpretable explanation. Our FKG structure encodes forensic traces along with their causal dependencies and links to scene content. To generate accurate FKGs, we introduce a novel forensic authentication network and an Iterative Context Refinement strategy that guides VLMs to produce faithful, grounded explanations. We also present FKG-50K, a dataset of 50,000 realistic forgeries with ground-truth FKGs. Experiments demonstrate that FKG outperforms both forensic detectors and VLMs in detection, forgery identification and localization, and forensic justification.

Researchers propose TheProfessor, a multi-teacher extension of PromptKD for distilling vision-language models. Using an ensemble of a domain-finetuned teacher and a zero-shot teacher, confidence-weighted ensembling improves harmonic mean accuracy by 1.77 points on average, with significant gains on domain-shifted datasets like EuroSAT.

REALM provides the first unified red-teaming benchmark for physical-world vision-language models, integrating 12 attack methods, 3 defenses, and 13 models to enable fair comparison of vulnerabilities. Key findings include the effectiveness of text and typographic attacks and the limited protective role of model scale.

A new method called HeRA aligns attention heads individually in multimodal LLMs, improving performance and reducing hallucinations.

Vision-Language Models (VLMs) are brittle to negation, relying on shallow co-occurrence and susceptible to misleading cues. HANCLIP restructures the embedding space with hyperbolic geometry and an angular triplet objective to encode negations, trained on 20K quadruplets, improving negation benchmarks without degrading standard performance.

ABACUS is a unified vision-language model that handles object counting, crowd counting, referring-expression counting, and count-faithful image generation without any benchmark-specific training. Built on a 3B-parameter foundation model, it introduces three key innovations: density-aware adaptive zooming with objectness maps, a boundary-aware count policy via GRPO, and a cycle-consistent GRPO strategy. It achieves state-of-the-art results across seven benchmarks, outperforming task-specific specialists and larger generalist models.