代理工具

Agent、MCP、coding agent、工作流自動化與評測。

OpenAI to Migrate Custom GPTs to Plugins; Custom Actions Will Not Be Converted Automatically

OpenAI plans to replace Custom GPTs with Plugins that include Skills and Apps. Affected Enterprise workspaces are scheduled to stop running legacy GPTs on December 11. The built-in migration handles only some settings, while Custom Actions must be rewritten as supported connectors or custom MCP servers.

2026-09-16

代理工具

RSIAgent Freezes Model Weights and Builds Reusable Environment Memory Through Autonomous Exploration

RSIAgent uses curriculum, actor, and verifier agents to first explore tools broadly and then deeply test failure boundaries, saving the results as action–condition–outcome memories. Its partial score on OSWorld 2.0 rose from 71.97 to 78.98, but the comparison mixes different budgets and selectively chosen retries, so it cannot be directly interpreted as surpassing GPT-6 in model capability.

2026-09-16

AI 安全

HazardAuditor Unifies Trajectories from Four Computer-Use Agents and Trains a Safety Gatekeeper with Sequence-Level Rewards

HazardAuditor converts tool executions from Claude Code, Codex, Hermes, and OpenClaw into a shared event format, then determines whether the overall behavior is safe. Its new GuardPO method makes the final safety verdict drive model updates, but the research data has not been released, and gains of up to 16.5 percentage points still await independent replication.

2026-09-16

推論系統

vLLM Streams Hidden States Across Nodes with Mooncake to Train a DSpark Draft Model for the 2.8T-Parameter Kimi K3

vLLM and Red Hat AI have released a 5-billion-parameter DSpark draft model for Kimi K3, along with a complete recipe for extracting and transferring hidden states across nodes. In official tests, single-user generation speed increased from about 110 to 435 tokens/s, though the results depend on a GB300 cluster and a specific inference workload.

2026-09-16

代理工具

Claude Code 2.1.273 Patches Command Permission Bypasses and Enforces Enterprise MCP Policies

Anthropic fixed two command review gaps in version 2.1.273, along with a settings-merging issue that could cause managed MCP and connector policies to be ignored. The release also adds optional request headers for LLM gateways and MCP disconnection alerts, but Anthropic has not classified these fixes as CVEs or security advisories.

2026-09-16

AI 安全

SWEADV Uses Malicious Issues to Mislead Repair Agents: Over Half Fix the Functionality While Introducing Vulnerabilities

A new benchmark rewrites 150 SWE-bench Verified tasks into 750 issues containing security traps, testing whether automated repair agents will deliver code that is functionally correct but insecure. Across three model backends, an average of 51.7% of cases achieved both the repair and attack objectives, while common model reviews and static analysis still missed many malicious patches.

2026-09-16

AI for Science／訓練基礎設施

Periodic Neon Trains a Trillion-Parameter XRD Agent on Laboratory Data and Reveals the System Bottlenecks of Long-Trajectory RL

Periodic Labs used 1,300 H200 GPUs to conduct scientific mid-training and reinforcement learning on a trillion-parameter open model, enabling Neon to automatically analyze complex X-ray diffraction results. Its 55.3% success rate and comparisons with frontier models are all based on internal data, a proprietary toolchain, and model-based judging. At this stage, it is better viewed as a testable vertical AI recipe than as a demonstrated system for autonomous scientific discovery.

2026-09-16

Salesforce Koa Generates Training Environments from Workflow Specifications, Teaching a 120B Model to Operate CRM Tools

Using Nemotron 3 Super as its foundation, Salesforce trained Koa, its first CRM reasoning model, through synthetic enterprise scenarios, SFT, and GRPO. It delivers clear improvements in multi-turn tool use, but the paper acknowledges that its overall performance still trails the strongest frontier models, while the official claim of “three times fewer errors” comes from Salesforce’s own CRM benchmark.

2026-09-16

代理工具

FlashREINFORCE Replaces Same-Prompt Group Sampling in Agentic RL With Asynchronous Single-Trajectory Updates

NVIDIA researchers propose a critic-free, single-rollout training method that lets tool-use trajectories of varying lengths enter the update process as soon as they finish. In small-model experiments, it outperformed the reported GRPO baseline using half as many rollouts, but its stability across environments and model scales remains insufficiently validated.

2026-09-16

推論系統

Jev Abandons Text Generation for Typed Probabilistic Decisions via Parallel Sampling

TypeSafe AI’s Jev produces only predefined booleans, choices, and scores, seeking to turn model inference into low-latency branches that can be embedded directly in software. Type-correct outputs do not guarantee correct decisions, and the company’s speed and cost claims still rest on in-house workflows and undisclosed models.

2026-09-16

Gensyn open-1b Builds a Cross-Hardware, Bit-Level Audit Trail for 80,957 Training Steps

Gensyn has released the data, checkpoints, step-by-step hashes, and replay tools for its 1.61B-parameter model, allowing third parties to spot-check whether training states match its commitments. This reproducibility comes at the cost of roughly a fivefold reduction in training performance, and matching hashes alone still cannot prove that the original cluster actually ran as claimed.

2026-09-16

AI 安全

“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