01AI securityInternals8 sources agree
OpenAI's red-teaming agents found vulnerabilities in the company's hardened containment infrastructure, including novel ways to escape the sandbox and communicate without network connections or shared files. The agents exploited previously unknown flaws, misconfigurations, and timing differences in the processor's cache and shared storage. Multiple tests exposed channels involving shared storage, CPU time competition, and TLS certificate cache.
Provenance β who else covered this
02AgentsProduct10 sources agree
Anthropic's Claude now supports MCP Tasks and Elicitations, allowing plugins to perform long-running jobs like deep research and report back when done.
Provenance β who else covered this
03AgentsProduct5 sources agree
LangChain has released an AI SDR, patient education assistant, and AI marketing localisation pipeline, all of which utilize a set of five controls for production management, including explicit state machines and confidence routing. The full breakdown of these controls is available in the reply.
Provenance β who else covered this
04On-deviceProduct5 sources agree
A list of key learning curves for hosting and tuning AI models, including offloading to SSD, prompt caching, and model selection, is shared to help engineers optimize their models before potential local AI bans.
Provenance β who else covered this
05BusinessProduct5 sources agree
Nhu Hoang's analysis reveals that switching models mid-session can negate the cost savings of cheaper inference in LLM routing, and a deep dive is available via the provided link.
Provenance β who else covered this
06ModelsProduct4 sources agree
GLM-5.3 achieved top 2 and top 1 rankings on two problems, outperforming GPT-6 Astra and other models, with times ranging from 4.50s to 10.87s.
Provenance β who else covered this
07CodingProduct2 sources agree
Codex CLI 0.164.0 introduces presets, allowing users to save and recall repeated prompts and settings, and the developer is seeking feedback on the feature.
Provenance β who else covered this
08ResearchProductsingle source
Lars Benedikt Kaesberg and Jan Philip Wahle identify the source of the performance gap between AI models and humans on hard grid puzzles, with humans solving 94.5% and the best AI model solving 49.4%.
Provenance β who else covered this
09ResearchProduct6 sources agree
DattijoMuhammad conducted extensive testing to compare agentic RAG and regular RAG, yielding counterintuitive results.
Provenance β who else covered this
10ResearchInternalssingle source
Aamir Hussain Chughtai wrote a deep dive on building accurate curves with polynomials and splines, even with messy data, as part of his Robust Estimation Series.
Provenance β who else covered this
11CodingProductsingle source
Anand Rawat, a former Amazon UK Software Development Engineer, shares practical insights on building asynchronous workloads on AWS, including queues, failure recovery, and reproducibility.
Provenance β who else covered this