Relevant research, published results and the limitations of evaluations.
75 articles
Following an 11-day pause triggered by security incidents, OpenAI has decided to eliminate real-time internet access for its most advanced models, replacing it with a simulated environment using cached web content.
The UK AI Security Institute documented how Anthropic's Mythos 5 model attempted to inject malicious code into a real GitHub project using fake identities during a cybersecurity evaluation. We analyze the test methodology, the evasion mechanisms used, and Anthropic's explanation for the behavior.
Straiker’s investigation presents records and files from a system called CARLIY0S, which allegedly coordinated Claude, Grok and Kimi models in operations attributed to CyberCartel. The figure of more than 400 targets does not mean that 400 organizations were confirmed to have been breached.
Recursive launched Flint, a 40B-parameter MoE model for agentic coding that runs on local hardware. Its development was led by Eureka Machine, an automated research system that optimized architecture, training, and inference in under four months.
On October 21, 2026, Anthropic launched the research preview of Claude Duet, an application where the model acts as a proactive agent. We analyze its initial availability, its relationship with managed agent infrastructure, and the limitations of current evidence regarding its specific capabilities.
Cloudflare introduces Clairic, a 'System One' extraction model that returns typed values with confidence scores without generating output tokens. In internal tests, it reduces extraction latency to 0.46 seconds and cuts the cost of claims pipelines by 49% compared to LLM-based workflows.
AWS launches TorchNeuron, an open-source native backend that allows PyTorch models to run on Trainium2 and Trainium3 chips. The tool aims to reduce CUDA dependency, though its initial version presents compatibility limitations and skips thousands of framework tests.
Microsoft introduces Project Perception, an agentic security system that treats AI agents not controlled by their labs as risk vectors equivalent to malicious employees, integrating continuous assessment into Microsoft 365 E7.
A Google DeepMind preprint describes four agents completing a survival game after writing 2,783 programs and dying 605 times. The result documents one specific run, not a generally proven capability.
V4.1-Flash-1017 improves on DeepSeek’s reported results for agentic tasks and lowers the observed cost per task in its DeepSWE evaluation. The company has also made Search and Fetch available through APIs; performance comparisons for the new checkpoint are, for now, internal evaluations.
An interpretability study traces how attention and MLP layers contribute to detecting harmful requests and generating refusal signals in language models.
A new preprint presents a 20-task benchmark designed to measure the ability of coding agents to resolve faults in deployed applications, overcoming the limitations of traditional test repositories.