1. MirroS Releases Code-as-World Open-Weight Models for Video-to-Physics Translation
MirroS has introduced Code-as-World, a framework that represents physical scenes as executable code rather than pixels or captions. Alongside the framework, they released Code-as-World-VL-4B and Code-as-World-VL-9B as Apache 2.0 open-weight models. Fine-tuned from Qwen3.5 and served via vLLM, the 9B model outperforms Gemini-3.1 Flash on the QuantiPhy benchmark. The system uses MuJoCo as an execution engine to verify physical consistency across five iterative rounds.
- • MirroS released Code-as-World-VL-4B and Code-as-World-VL-9B under the Apache 2.0 license.
- • The models are fine-tuned from Qwen3.5 and served as BF16 safetensors via vLLM.
- • Code-as-World-VL-9B scored 55.4 on the QuantiPhy benchmark, outperforming Gemini-3.1 Flash's score of 54.8.
- • The framework uses an agentic loop to propose, instantiate, execute, render, and verify physical consistency using MuJoCo.
- • The current implementation is limited to rigid-body physics and does not embed the discovery loop inside the model itself.
Developers can extract structured, executable physical world representations from video files using Apache 2.0 licensed models.
2. Claude Code Appends Session URLs to Git Commits by Default
Developers have discovered that Claude Code automatically appends session URLs to all git commit messages and pull request descriptions by default. This behavior occurs without an explicit opt-in prompt or warning during onboarding. To prevent these URLs from cluttering git history or exposing session details, developers can manually disable the feature by modifying the attribution.commit setting in their .claude/settings.json file.
- • Claude Code automatically appends a session URL to every commit message and PR description without an opt-in prompt.
- • Users report that the automatic URL inclusion clutters git history and appears unprofessional to external contributors.
- • The behavior can be suppressed by setting 'attribution.commit' to false in the '.claude/settings.json' file.
- • The setting is currently not easily discoverable during onboarding or first-time use.
Developers must manually disable this setting to prevent internal session URLs from being exposed in public git histories.
3. Anthropic Resets Claude Sessions Following Infostealer Malware Attacks
Anthropic has taken protective action after discovering that infostealer malware on some users' PCs successfully hijacked active Claude login sessions. To mitigate the impact of unauthorized usage and billing drains, Anthropic signed out the affected users, removed saved payment methods, and issued refunds. Developers are advised to ensure their local development environments are secure against session-stealing malware.
- • Anthropic issued a warning to Claude users regarding local infostealer malware.
- • The malware successfully hijacked active Claude login sessions from users' PCs to drain usage.
- • Anthropic responded by signing out affected users, removing saved payment methods, and issuing refunds.
Developers using Claude must secure their local environments to prevent session hijacking that could lead to unauthorized API usage and billing drains.
4. Google Releases EnvHarness to Adapt Static Agent Benchmarks for Training
Google Cloud AI Research, in collaboration with academic partners, has released EnvHarness under an Apache-2.0 license. EnvHarness is a programmable layer designed to adapt static agent benchmarks for policy training without modifying the underlying simulators. It uses an LLM-based designer called EnvRigger to diagnose policy flaws and generate targeted wrappers. Across five benchmarks, including SWE-bench Verified, the tool improved agent performance by up to 9.0 points and reduced execution steps by nearly 10%.
- • EnvHarness wraps existing environments using plug-in components through a standard step/reset interface.
- • The tool uses EnvRigger, an LLM-based designer, to diagnose policy flaws and write targeted wrappers.
- • It provides three modification components: Stage (replays actions), Contract (installs per-step hooks), and Chain (composes environments).
- • The tool is released under an Apache-2.0 license but excludes live user accounts or physical robots.
- • Testing on SWE-bench Verified and ALFWorld showed up to 9.0 point improvements and a 9.8% reduction in execution steps.
Developers can automatically diagnose policy flaws and wrap existing agent environments to improve performance.
5. Benchmark Shows Gemma 4 and Qwen Open VLMs Match Gemini Flash on Egocentric Data
Hebbian-Robotics has published an evaluation of open-weights vision-language models (VLMs) processing egocentric data using the HFlow tool. Testing on the Egocentric-10k dataset revealed that Gemma 4 26B-A4B and Qwen 3.8 27B perform nearly on par with a Gemini 2.5 Flash baseline. Notably, Gemma 4 achieved comparable results while being 19 times cheaper, making self-hosted open VLMs highly practical for private, high-throughput data processing.
- • Hebbian-Robotics evaluated open-weights VLMs on the Egocentric-10k dataset using the HFlow tool.
- • Gemma 4 26B-A4B achieved a 90.87% agreement with the Gemini 2.5 Flash baseline (91.65%) while being 19 times cheaper.
- • Qwen 3.8 27B scored 90.79% and GLM 5.3 Flash scored 91.00% in the same evaluation.
- • The study highlights cost, throughput, output reliability, and ease of self-hosting as the primary differentiators for modern open VLMs.
Developers can self-host Gemma 4 or Qwen 3.8 for private, high-throughput video and image processing at 19x lower cost than commercial APIs.