Xiamen University et al. open ProgramGym: 1,648 verifiable C/C++/Go/Rust environments built from GitHub repos for agents that rebuild programs from black-box behavior, plus a role-isolated Program-Agent harness. Fine-tuning Qwen3.8-27B on its trajectories raised ProgramBench from 34.87% to 42.84%; with Program-Agent the SFT model hit 56.25%.
Madhusudhan Vasanth Kumar’s Apache 2.0 terminal, VS Code, and Zed agent runs commands in Seatbelt (macOS) or bubblewrap (Linux) so they write only in the project, temp folders, and package caches, with home secrets unreadable and no network by default; edits show a diff first, and a hash chained audit log records every permission decision. Install with npm install g @garuda agent/garuda ; it works with Claude, OpenRouter, and local OpenAI compatible models…
Penn State, Montpellier, and Dartmouth showed LangGraph, AutoGen, Google ADK 2.0, AWS Strands, Microsoft Agent Framework, and CrewAI will execute a structured tool call if the model is prompted to echo it—even when the model later refuses. Their ToFu fuzzer also confirmed 253 bugs across 118 open source MCP servers, including 64 path traversal and three command injection flaws…
NVD published CVE 2026 108953 on Oct 12 for cfdude's mac shell mcp: before 2.0.0, clients (or malicious prompts) can call a whitelisted command like pwd with a semicolon in the arguments and run forbidden shell as the server user without approval. Upgrade to 2.0.0 or later…
After you connect a GitHub repo and enable code review, comment @codex review (or turn on Automatic review) and Codex posts a P0/P1 review that follows AGENTS.md Code Review Rules; research preview Security Review uses @codex security review . You can then ask @codex fix the P1 issue to push a follow up fix to the branch when it has permission…
Subquadratic's public preview Snap scores bounded answers (yes/no, choice, score) over up to 384K tokens of text and images in about a second at 100K, so coding agents can offload file selection, tool choice, and test verification while leaving generation to a frontier LLM. It averages 67.6 vs Jev's 61.6 on 16 enterprise Decision Index tasks, costs $0.05 per million input tokens with free outputs, and uses Subquadratic Sparse Attention instead of quadratic prefill…
Spotted something we missed? Start a thread.
