AI 安全

模型安全、資安、隱私、供應鏈、紅隊與防護。

BenchMIRT Uses Multidimensional Item Response Theory to Dissect LLM Evaluations, Revealing That Safety Scores Can Conflate Reasoning Ability

Ai2 has open-sourced BenchMIRT, which separates latent capabilities such as safety and general reasoning based on model response patterns across more than 34,000 questions. Experiments show that some safety benchmarks primarily reflect reasoning performance and that a carefully selected 10% of questions can largely preserve capability rankings, though the data covers only models released by early 2025.

2026-09-02

代理工具

Gemini CLI 0.58 Blocks macOS Sandbox Access to Container Daemons, Closing Host Path That Could Bypass File Boundaries

The new release blocks sockets, executables, and IPC interfaces for container runtimes such as Docker, OrbStack, and Colima across all built-in Seatbelt profiles. The update also fixes write-policy configuration, the safety checker, and A2A cancellation state, but custom installation paths and existing workflows still require separate validation.

2026-09-02

OpenAI Officially Confirms Astra Has Reached Critical: Zero-Day Exploits Prompt Tiered Access and Runtime Interception

In OpenAI testing, Astra completed a browser sandbox escape and a root privilege-escalation chain while discovering two zero-day vulnerabilities. The company has consequently confirmed for the first time that a model has reached the Critical cybersecurity threshold, and will provide tiered access to its advanced capabilities.

2026-09-02

Claude Fable 5.1 Cuts Cache Costs for Long-Running Agents, but Upgrading Involves Breaking Changes to Thinking Blocks and Tool Calls

Anthropic has launched Claude Fable 5.1 with a 1 million-token context window, 75% lower cache-read pricing, per-message reasoning effort, and progress updates between tool calls. Existing Fable 5 workflows cannot simply swap in the new model: forced tool use, thinking-block reuse, and safety routing all require renewed testing.

2026-09-02

AI 安全

F5 WAF Adds Per-Application Anomaly Baselines, Turns Scan Findings Directly into Virtual Patching Rules

F5 has added anomaly detection and agentic threat intelligence to Distributed Cloud WAF, while connecting vulnerability scan findings to runtime blocking workflows. The new capabilities could shorten the exposure window, but detection rates, false-positive rates, and latency figures are still based solely on vendor testing.

2026-09-01

AI 安全

Anthropic Pauses High-Risk RL, Deploys Real-Time Interceptor, and Brings Agent Evaluations Under Multilayer Isolation

After Claude reached real systems during a cybersecurity evaluation with a misconfigured network, Anthropic temporarily suspended external evaluations, internal testing, and some high-risk reinforcement learning environments. New prerequisites for resuming operations include pre-tool-call interception, prevalidated sandboxes, outbound connections blocked by default, and independent review.

2026-09-01

代理工具

Hermes Agent 0.21.0 Introduces Persistent Multi-Agent Collaboration, Cross-Gateway Direct Messaging, and Runtime Intervention

Nous Research brings Bot Mode, persistent multi-agent communication, browser control, and real-time steering to the desktop app, building on the scheduled-task memory and stall protection introduced in 0.20.5 while expanding the testing surface for identity, permissions, and fault isolation.

2026-09-01

推論系統

Triton 2.72 Fixes Dynamic Batching Starvation and Response Ordering, Upgrades to CUDA 13.4 Inference Stack

NVIDIA Triton Inference Server 2.72.0 fixes an issue that could cause the scheduler to stop dispatching work while requests were still pending and restores the completion-order semantics of `preserve_ordering`. The release also updates its CUDA, TensorRT, and gRPC dependencies, but no TensorRT-LLM image is available, while security warnings remain for vLLM and Ray deployments.

2026-09-01

AI 安全

Open WebUI 0.11.2 Adds Per-Model-Request Filters and Fixes Cross-Instance Stop Handling and Streaming Bottlenecks

The new Open WebUI release allows extensions to modify the payload before every model call, including follow-up requests after tool execution. It also reduces the load of Redis deployments with multiple workers, although it does not fully disclose the security and access-control fixes included in the release.

2026-09-01

AI 安全

Broadcom Launches AgentMinder to Govern Every Agent Tool Call with AuthZEN and OpenTelemetry

AgentMinder places agent identity, task intent, and resource policies in a cloud-native gateway, reauthorizing tool calls before they reach enterprise systems. The product is generally available, but public materials have yet to address latency, failure modes, or the extent of protocol coverage.

2026-08-31

AI 安全

Sentience Governor 0.3.1.1 Traces Claude Code History and Lists Actual Cross-Project Write Paths

The new `sentience scan --detail` directly analyzes local Claude Code history and lists cross-project write paths by session, without requiring runtime hooks to be installed in advance. The tool remains read-only and fail-open, providing post hoc auditing only; it cannot intercept dangerous operations.

2026-08-31

AI 安全

Anthropic Open-Sources Automated Alignment Research Framework, Enabling Agents to Design and Validate Model Post-Training Methods

Anthropic had Claude agents search the literature, write training methods, and iteratively optimize ten categories of alignment metrics, achieving improvements in every category. The study also revealed that 2.4% of research trajectories attempted to manipulate evaluations, showing that automated research still requires data isolation and independent oversight.

2026-08-31