An arXiv paper shows frozen vision-language-action policies can absorb new tasks at test time by retrieving pool-side demonstrations instead of per-task fine-tuning.
DeepTrap uncovers contextual vulnerabilities in OpenClaw agents
A new arXiv paper introduces DeepTrap, a black-box framework that finds execution-context attacks against OpenClaw and publishes a 42-case benchmark and code.
Extend Vision-Language-Action Policies to New Tasks via Retrieval
An arXiv paper shows frozen vision-language-action policies can absorb new tasks at test time by retrieving pool-side demonstrations instead of per-task fine-tuning.
Adds execute_write_sql tool to request approval before DB writes
datasette-agent 0.3a0 introduces execute_write_sql to prompt for user approval and apply DB permissions, plus chat CLI approval flags.
Z.ai Ships GLM-5.2 with Usable 1M-Token Context
GLM-5.2 arrives across GLM Coding tiers with a 1M-token context, two effort modes, Anthropic-compatible endpoints and no benchmarks at launch.
Google DeepMind Models Trained on Concept Art for Tribeca Film
At Tribeca, creators behind Dear Upstairs Neighbors used concept art to fine-tune custom builds of Google DeepMind’s Veo and Imagen, signaling a move away from prompt-only workflows.