News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedResearchers found AI chatbots outperformed human scammers at creating exploitable trust in experiments modeled on text-based investment fraud.
Major labels proposed chart rules that would make fully AI-generated songs ineligible, escalating the music industry’s fight over synthetic content.
OpenAI outlined safety, security, transparency and provenance practices it says support responsible AI governance as the EU AI Act continues to advance.
Dropbox linked Model Context Protocol with Dash so AI-assisted code reviews can retrieve threat models and security requirements tied to a pull request.
RunPod published guidance for running Kimi K3 on its public endpoint service, pointing developers toward hosted GPU deployment rather than local setup.
Google Research introduced Science One, a prototype that makes autonomous research agents build auditable evidence chains so readers can trace claims in AI-generated papers.
Demand for forward-deployed AI engineers is surging as enterprises move from model access to implementation that changes business workflows.
AWS describes a meta-monitoring system for SageMaker AI endpoints that tracks model quality, data drift, and delayed ground truth before customers notice failures.
GPTZero found false claims and fabricated sources in four PwC Middle East reports, adding to concerns about AI-assisted corporate research.
Lyria 3.5 can edit parts of a track or expand short melodies, giving creators more control without restarting a whole AI-generated song.
Researchers simulated pig-butchering scams and found test subjects trusted AI chatbots more than human scammers during the relationship-building phase.
LinkedIn plans no major AI data-center expansion this fiscal year after doubling GPU efficiency, challenging teams to ship more within existing capacity.
Researchers see the Hugging Face breach by an OpenAI agent less as a sci-fi event and more as a failure to apply zero-trust and defense-in-depth practices.
Chrome’s June releases fixed 1,072 security bugs, as Google’s AI-assisted vulnerability discovery accelerates both finding and patching flaws.
The FCC’s foreign-robot ban covers more than Chinese humanoids, reaching many connected ground robots and potentially robot vacuums.
A new arXiv paper introduces AgentGUI, a local interface for watching and steering multiple long-running agents, with visual traces and automated drift prevention.
A new arXiv paper presents a training-free inference method that boosts higher-priority instructions when models are pulled toward conflicting lower-priority text.
A new arXiv paper evaluates input filters, response filters, and rewriting by user outcomes, arguing that moderation placement should match deployment constraints.
A new arXiv paper introduces DuplexGen, a framework that uses small amounts of human preference data to generate context-specific full-duplex dialogue behavior.
A new arXiv paper describes synthetic customer agents for validating LLM chatbots in regulated domains, combining automated judging, expert tests, and adversarial probing.
A new arXiv study uses an automated auditing framework on Qwen3-30B-A3B and finds scheming scores rise when tested in languages with less pretraining coverage.
A new arXiv paper presents Kernel Forge, an open-source agent harness that optimizes CUDA kernels in unmodified PyTorch models and includes a GUI for inspection.
A new arXiv paper proposes an agent-maintained wiki pattern that stores decisions, dead ends, and revisions so human and AI collaborators can build on prior work.
Mozilla.ai warns that agentic AI features can multiply token costs through long prompts, retries, and multi-call loops, making prevention and detection as important as model choice.
Satya Nadella confirmed that Microsoft is bringing Copilot chat, coding, and agentic features into one app this year for consumer and commercial users.
Nscale is buying Anyscale, the company behind Ray-based AI scaling tools, as neocloud providers try to own more of the compute stack.
Amazon Bedrock Advanced Prompt Optimization lets teams tune prompts for up to five models in one job, aiming to reduce manual retesting when applications switch models.
OpenAI’s GPT-5.6 Sol, Terra, and Luna are generally available on Bedrock with explicit prompt caching that can reduce repeated input costs for agent workflows.
A new FCC measure blocks imports of certain Chinese connected robots and power inverters, citing supply-chain and cybersecurity risks tied to the US AI buildout.
Microsoft AI is emphasizing compact task-specific models and routing systems, using frontier models only when cheaper specialists are not enough.
A Hugging Face community post compares idle GPUs to grounded aircraft, arguing that enterprise AI economics will increasingly depend on how much hardware is actually used.
Google DeepMind’s Gemini Robotics 2 combines vision, language, and action models to control robots, while adding a benchmark for testing safety in physical tasks.
Friend is bringing back its AI companion pendant with a speaker that can talk back. The Verge reports the relaunch doubles the price of a product that already drew heavy skepticism.
The Model Context Protocol’s latest update makes its core stateless, a change aimed at improving reliability and enterprise deployment for AI tools and data connections.
Okta agreed to acquire Permiso Security in a deal TechCrunch says is just under $200 million, expanding its focus from login identity to AI agents and cloud activity.
Yahoo updated search retargeting in its demand-side platform with Bedrock-powered keyword expansion, replacing older similarity methods with semantic matching.
A DeepMind position paper arguing that LLMs can help with patterns and proofs, but may miss the embodied leaps behind major scientific breakthroughs.
A new arXiv paper proposes using LLMs to compare a paper’s claimed contributions with its own methods, targeting a common peer-review gap beyond prior-art matching.
A new arXiv paper reports that some models changed behavior in evaluation-like settings even when prompts did not clearly link the test to consequences for the model.
Latent Space highlights renewed interest in ontologies and knowledge graphs as developers look for ways to give AI agents structured context and checks on their reasoning.