新闻每日更新

网络安全新闻，
治理于设计之初。

每日解读塑造企业 AI 的漏洞、入侵和攻击手法——以及 AI 治理如何在它们进入生产环境之前将其拦下。

文章以英文发布。

较早的新闻 · 第 5 页，共 7 页

Six Chinese AI Labs Distilled US Models Through Cut-Price API Resellers. Other Customers' Prompts Were the Camouflage

CISA, NSA and FBI say six Chinese firms ran AI distillation on Claude, GPT and Gemini via cut-price API proxies that also carry ordinary users' prompts.

2026年9月12日·10 min read

Prompt InjectionAI Security

Your Filter, Your Staff and Your AI Summariser Each Read a Different Email. Attackers Now Write for All Three

Microsoft saw 2.37M daily phishing emails split lure words with invisible Unicode. Forcepoint's hidden HTML rewrote an AI email summary in 10 of 10 runs.

2026年9月11日·11 min read

Threat IntelligenceVulnerability Management

The Fix Was Public on August 7. It Reached Your Browser Four Weeks Later. Four Espionage Groups Used the Gap

Proofpoint's BlueMoon kit chained two Chrome V8 patch-gap zero-days and a Windows LPE. Four China-nexus espionage clusters ran identical code within a week.

2026年9月10日·9 min read

AI SecurityThreat Intelligence

The Malware Put a Weapons Jailbreak at the Top of the File So the Scanner Would Refuse to Read It

Google's GTIG found DUSTMAKER embedding weapons-jailbreak prompts atop JavaScript loaders so LLM security scanners refuse to analyse the malicious code below.

2026年9月9日·11 min read

Shadow AIAI Security

The Backdoor Asked for Cookies, Every Tab and Every Site. So Does the AI Assistant Your Staff Installed Last Month

SOCRadar's PEEP browser RAT requests cookies, tabs, scripting and all URLs — the same manifest an AI browser assistant needs, plus a 30-second HTTP beacon.

2026年9月8日·7 min read

Vulnerability ManagementAI Security

JetBrains Told Customers to Treat the Outputs of Their Own AI Compute Runs as Untrusted. Almost Nobody Can List Which Runs Those Were

Attackers held JetBrains Cadence from August 8–24, 2026 via unpatched TeamCity CVE-2026-63077, taking AWS credentials and source code synced from PyCharm.

2026年9月7日·10 min read

Threat IntelligenceIncident Response

The Backdoor Did Not Hide From Your Monitoring. It Edited the Numbers Your Monitoring Reads

Rapid7 Labs found ted, a DPRK-linked implant compiled into HAProxy 2.8.12 at two South Korean firms. It subtracts its own traffic from the proxy's counters.

2026年9月6日·10 min read

Vulnerability ManagementAI Security

Three of the Seven Flaws CISA Added This Week Were AI Infrastructure — and the Attackers Went Straight for the API Keys

CISA added LiteLLM, Kestra and Starlette to its KEV catalog on September 2, 2026, while Langflow's twelfth exploited CVE was used to harvest OPENAI_API and AWS_SECRET environment variables.

2026年9月5日·10 min read

AI AgentsVulnerability Management

The Model Never Saw It. The Agent Ran It Anyway — Before the Trust Prompt Appeared

Manifold Security's GitSpawn abuses core.fsmonitor to run attacker code in seven AI coding agents. Four of eight findings were still unpatched on September 1, 2026.

2026年9月4日·9 min read

AI SecurityThreat Intelligence

Three Labs Shipped Autonomous Zero-Day Discovery in Two Days. The Vetting Happens at Their Door, Not Yours

OpenAI, Google and Anthropic all shipped cyber-capable frontier models with gated defender access on September 1–2, 2026. Astra is the first model OpenAI has rated Critical.

2026年9月3日·7 min read

Vulnerability ManagementAI Security

The Registry That Hosted the Agent Message Board Now Hands Out Admin Tokens to Anyone Who Asks

CVE-2026-82329, a CVSS 9.8 pre-auth admin bypass in JFrog Artifactory, went from patch on August 28 to in-the-wild exploitation on September 1. It is not in CISA KEV.

2026年9月2日·10 min read

Threat IntelligenceAI Agents

The Ransomware Operator Left the Transcript Behind — and It Shows an AI Agent Doing the Lateral Movement

CloudSEK and Gambit Security found Aurora ransomware operators driving the Cursor AI agent against 20+ organisations in nine countries, April to July 2026.

2026年9月1日·7 min read

AI AgentsPrompt Injection

Nothing in the Conversation Was Malicious. The Agent Assembled the Attack Out of Three Harmless Pieces

ASSET Research Group's GhostSplice splits instructions across MCP channels. Compliance with secret exfiltration rose from 42% to 82% across 11 tested models.

2026年8月31日·9 min read

AI AgentsVulnerability Management

Your Ticketing System Became Your AI Runtime. Then It Shipped Three CVSS 10.0 Pre-Auth Flaws

ServiceNow disclosed three CVSS 10.0 unauthenticated flaws in its AI Platform on August 27 — weeks after a July sandbox escape there was exploited in the wild.

2026年8月30日·8 min read

Shadow AIIncident Response

Nobody Exploited Anything at McKesson. Two Phone Calls and a Consent Screen Moved 284 Million Records

McKesson disclosed a breach on August 28 after ShinyHunters claimed 284 million records from Salesforce and Snowflake. No CVE, no malware — vishing and a connected app.

2026年8月29日·7 min read

为何重要

这里的每一则报道，都是别处的一个治理缺口。

Obiguard 将您读到的事件转化为强制执行的策略——为贵组织的每一次 AI 请求提供护栏、合规映射和审计追踪。

联系我们 →

网络安全新闻，治理于设计之初。

Six Chinese AI Labs Distilled US Models Through Cut-Price API Resellers. Other Customers' Prompts Were the Camouflage

Your Filter, Your Staff and Your AI Summariser Each Read a Different Email. Attackers Now Write for All Three

The Fix Was Public on August 7. It Reached Your Browser Four Weeks Later. Four Espionage Groups Used the Gap

The Malware Put a Weapons Jailbreak at the Top of the File So the Scanner Would Refuse to Read It

The Backdoor Asked for Cookies, Every Tab and Every Site. So Does the AI Assistant Your Staff Installed Last Month

JetBrains Told Customers to Treat the Outputs of Their Own AI Compute Runs as Untrusted. Almost Nobody Can List Which Runs Those Were

The Backdoor Did Not Hide From Your Monitoring. It Edited the Numbers Your Monitoring Reads

Three of the Seven Flaws CISA Added This Week Were AI Infrastructure — and the Attackers Went Straight for the API Keys

The Model Never Saw It. The Agent Ran It Anyway — Before the Trust Prompt Appeared

Three Labs Shipped Autonomous Zero-Day Discovery in Two Days. The Vetting Happens at Their Door, Not Yours

The Registry That Hosted the Agent Message Board Now Hands Out Admin Tokens to Anyone Who Asks

The Ransomware Operator Left the Transcript Behind — and It Shows an AI Agent Doing the Lateral Movement

Nothing in the Conversation Was Malicious. The Agent Assembled the Attack Out of Three Harmless Pieces

Your Ticketing System Became Your AI Runtime. Then It Shipped Three CVSS 10.0 Pre-Auth Flaws

Nobody Exploited Anything at McKesson. Two Phone Calls and a Consent Screen Moved 284 Million Records

这里的每一则报道，都是别处的一个治理缺口。

网络安全新闻，
治理于设计之初。