AI
8 items · 17 articles in this topic
Model releases came with safety caveats, and Anthropic's IPO filing is generating a stream of risk and governance coverage. A claimed result from frontier labs on a complexity conjecture is the most technically notable item.
- 1
OpenAI says GPT-6.1 Terra is safer overall than older models but showed more destructive-action flags in safety testing.
- 2
Frontier AI labs claim their models refuted the Exponential Time Hypothesis, a major complexity-theory conjecture. The summary gives no detail on verification.
- 3
Google DeepMind released Genie 4, a world model adding real-time audio, more realistic physics and humans, and mid-simulation prompts. It is in research preview for trusted testers, developers and AI Ultra subscribers.
- 4
Anthropic's IPO risk factors list seven potential AI risks, including blackmail and loss of control. The prospectus runs 335 pages, with about 100 on risks.
- 5
Mathematicians criticize OpenAI's new approach to sharing research breakthroughs after earlier preprint releases.
- 6
Anthropic is expanding its GSA partnership to make Claude available across all levels of US government. It also published early research on using Claude to accelerate drug development.
- 7
Stanford Daily reports Anthropic allegedly trained Claude on pirated copies of Stanford professors' books.
- 8
Apple is adding an AI assistant to Pages, Keynote and Numbers, built on Apple and OpenAI models, with MCP support on macOS 27. Opera proposed WebAI, a standard for running AI models locally in the browser.
More in AI · 9 articles›
DeepMind unveiled Genie 4, a world model promising better physics and humans, rolling out to developers and AI Ultra subscribers.
Ars Technicagenieworld-modelgoogle-deepmindFeed excerpt
Genie 4 will roll out to developers and Google AI Ultra subscribers soon.
AgentSentry proposes temporal causal diagnostics and context purification to defend LLM agents against multi-turn indirect prompt injection attacks.
arXiv cs.CR (Cryptography & Security)llmprompt-injectionagentsFeed excerpt
arXiv:2602.22724v2 Announce Type: replace Abstract: Tool-augmented large language model (LLM) agents aggregate dialogue history, tool outputs, and retrieved content into an accumulated runtime state to complete multi-step tasks. This execution model exposes agents to indirect prompt injection (IPI) attacks, in which attacker-controlled instructions embedded in external artifacts enter the agent state and later influence tool use. Multi-turn IPI is especially challenging because injected content can persist across steps and trigger unauthorized state-changing actions only after merging with user intent. This creates a source-attribution problem: existing inference-time defenses can often detect suspicious behavior but do not directly distinguish whether a pending action is supported by the user task or influenced by injected content. They therefore resort to conservative blocking, termination, or broad filtering. We propose AgentSentry, an inference-time defense that operationalizes temporal causal diagnostics through boundary-local counterfactual re-executions at tool-return boundaries. It substitutes the mediator under a restored snapshot, records whether the pending proposal surv…
Anthropic published early research results on using Claude to accelerate drug development.
Hacker News (front page)claudellmbiotechFeed excerpt
Article URL: https://www.anthropic.com/research/early-progress-drug-development Comments URL: https://news.ycombinator.com/item?id=50265609 Points: 13 # Comments: 1
Liquid AI introduced LFM3, a model family it says covers every modality for specialized intelligence.
Hacker News (front page)liquid-aimultimodalllmFeed excerpt
Article URL: https://www.liquid.ai/blog/lfm3 Comments URL: https://news.ycombinator.com/item?id=50262229 Points: 17 # Comments: 6
Gizmodo reports police are using AI tools to pull over drivers for behavior deemed suspicious.
Hacker News (front page)ai-surveillancepolicingprivacyFeed excerpt
Article URL: https://gizmodo.com/police-are-using-ai-to-pull-over-drivers-for-suspicious-behavior-2000831220 Comments URL: https://news.ycombinator.com/item?id=50261819 Points: 130 # Comments: 55
Anthropic's privacy page explains that submitting chat feedback can make conversations usable for model training.
Hacker News (front page)llmprivacyanthropicFeed excerpt
Article URL: https://privacy.claude.com/en/articles/7996868-is-my-data-used-for-model-training Comments URL: https://news.ycombinator.com/item?id=50259406 Points: 56 # Comments: 28
Opera proposed WebAI, a new web standard for running a variety of AI models locally within the browser.
TechCrunchbrowserlocal-aiweb-standardFeed excerpt
The new standard, WebAI, is designed to work with a variety of AI models.
IEEE Spectrum reports on Wikipedia's WikiProject effort to identify and clean up AI-generated articles and writing.
Hacker News (front page)wikipediaai-contenteditingFeed excerpt
Article URL: https://spectrum.ieee.org/wikipedia-wikiproject-ai-writing-cleanup Comments URL: https://news.ycombinator.com/item?id=50257326 Points: 33 # Comments: 12
An investigation finds UK ministers met AI datacenter executives frequently while environmental groups were largely excluded from AI Growth Zone decisions.
The Registeruk-policydatacentersai-policyFeed excerpt
Investigation finds businesses dominated govt meetings while communities were left out of decisions about AI Growth Zones