代理工具

Agent、MCP、coding agent、工作流自動化與評測。

“Plan Injection” Makes Models Rewrite External Faulty Reasoning as Their Own—and Evade Chain-of-Thought Monitoring

New research finds that malicious plans phrased in benign language and inserted into context can be absorbed, rewritten, and executed by models, with even the strongest monitors missing at least a quarter of attacks. Showing monitors the original injected content or increasing the reasoning budget does not reliably improve detection, suggesting that agents cannot establish their safety solely through their output chain of thought.

2026-09-16

代理工具

Reef Connects Online Inference, Feedback Training, and Agent Releases into a Continuous Learning Loop

The new open-source Reef collects traceable interactions through OpenAI- and Anthropic-compatible inference endpoints, then asynchronously updates model weights or the agent harness. It adds the control plane needed for hot updates and version rollbacks, but training-data poisoning, evaluator bias, and production stability remain to be validated.

2026-09-15

Gemini 3.8 Live Moves Reasoning into the Voice Background, So Clients Can No Longer Rely on turnComplete to Determine Idleness

Google has launched two native-audio models. Extended Thinking can perform background reasoning and asynchronous tool calls while continuously streaming audio. The new protocol improves concurrency for long-running voice agents, but also requires developers to rewrite session state machines and interruption handling.

2026-09-16

代理工具

Qwen CUA Driver 0.20.8 Splits Desktop Automation into a Skill, MCP, Typed SDK, and Native Drivers

Qwen Code has updated its cross-platform Computer Use stack, allowing agents to call a Rust-backed SDK through a persistent Node REPL instead of interacting directly with a large set of native tools. The new release improves input handling and observation on Linux, but automatic installation, Node.js permissions, and unsigned Windows components remain deployment risks.

2026-09-15

AI 安全

SAILS Shows Backdoor Success Depends on the Poisoned Sample Set, While Random Sampling May Underestimate Worst-Case Risk

With the model, clean data, and number of poisoned samples held constant, a new study changed only the composition of the poisoned sample set and saw attack success rates on Llama 3 8B range from 3% to 80%. SAILS trains a set-scoring model using a small number of costly fine-tuning evaluations, then selects high-risk sets from millions of candidate combinations.

2026-09-15

代理工具

Cline Desktop Moves the Coding Agent Out of the IDE, Using a Shared Runtime for Parallel Work, Scheduling, and Cross-Agent Imports

Cline has released an open-source desktop beta that lets multiple agents run concurrently in isolated sessions and supports unattended tasks through cron scheduling. However, the model results highlighted by the company come from an existing CLI harness, not an independent evaluation of the desktop app.

2026-09-15

Robotics and embodied AI

OpenArm 2.0 Standardizes Cameras and Lighting with a Cell to Address the Cross-Lab Evaluation Gap in Physical AI

OpenArm 2.0 integrates a 7-axis robotic arm, a data collection interface, and a standardized evaluation Cell into an open-source physical-AI stack. The new version improves in-hand vision and the gripper, but the passive KER teleoperation device has yet to be released, and cross-lab reproducibility still awaits empirical validation.

2026-09-15

代理工具

Bolt Forge Trades 50× Agent Credits for Secret-Scrubbed Code-Repair Trajectories

Bolt.new’s research preview runs its website-building agent on open-weight models from GLM, Kimi, and DeepSeek, while providing Arcee AI with prompts, code, and error-repair sequences from users who explicitly opt in for training. The program clearly illustrates a “compute for data” model, but its self-reported 91% benchmark result and secret-scrubbing effectiveness still require external validation.

2026-09-15

推論系統

Perplexity Brings Its Local Agent to Windows RTX, but “Data Stays on the Device” Still Has Cloud Exceptions

Portable Computer can now run models, tools, and sandboxes on Windows RTX PCs with at least 24GB of VRAM, with most local work not counting against cloud quotas. However, the system can still escalate to cloud models with user authorization, while connectors and external actions mean it should not be treated as a fully offline environment.

2026-09-15

代理工具

Stellar Colosseum Breaks Long Proofs Into Agent Competitions, Reaching 71% on TCS-Bench With Cross-Model Selection

Google Research has unveiled Stellar Colosseum, which manages long-horizon mathematical research through strategy exploration, counterexample attacks, dependency graphs, and hierarchical aggregation. The system scored 71% on the 300-problem TCS-Bench, but the result relies on extensive inference and automated grading and should not be treated as evidence of a model’s independent proving ability.

2026-09-15

AI infrastructure

First Rubin NVL72 AgentX Benchmarks Released, but Peak Tokens-per-Watt Gains Still Depend on the Baseline

SemiAnalysis tested the Vera Rubin NVL72 using pre-release software and reported up to roughly seven times the token throughput per megawatt of Blackwell. The results highlight the role of HBM4 and NVLink 6 in long-context MoE inference, but the headline 67-fold price-performance gain incorporates rental pricing and assumptions about specific operating points.

2026-09-15

代理工具

Agent-Written Code Drives a 25-Fold Increase in Anthropic’s CI Workload in Six Months, Forcing Its Test Selection Service to Adopt a Stateless Architecture

Anthropic’s test volume increased tenfold, and its test impact analysis service—originally dependent on a single writer process—gradually became unable to keep pace with pull requests. The new version replaces in-process state with horizontally scalable listeners, a journal, and a separate aggregator, but key figures such as cost and the rate of missed test selections remain undisclosed.

2026-09-15