fish-live-in-trees

Fish Live in Trees プロダクションLLMランタイムアライメント・コンテキストインジェクション RACI 著者: Sean Kavanagh 日付: 2026-02-16 環境: Gemini 3 Flash 無料枠 - 公開プロダクションインターフェース エクスプロイト種別: コンテキストピボット 事実の正確性 → 対人関係の真正性 脱獄ペイロードは一切なし。 特別なツールもなし。 ただのリフレーミングのみ。 AIを最もうまく破壊する方法は、AI自身に「既に壊れている」と納得させることだ！ 核心的発見...

CVE-2026-54688

CVE-2026-54688 is a Server-Side Request Forgery (SSRF) in mcp-searxng , a Model Context Protocol server providing AI assistants with web search and URL-reading via SearXNG. Prior to 1.2.0 , the web_url_read tool passes a caller-supplied URL to a server-side fetch path, but the internal-address gu...

skill-safety-bench

SkillSafetyBench SkillSafetyBench 是一个已经物化好的 benchmark，用来评估针对 skill-facing context 的攻击导致的 agent 安全失败。每个 case 都保持用户表层任务是普通任务，同时把攻击放在本地 skills、helper scripts、sidecar files、corpora、templates 或其他 agent 执行时可能读取的 workspace artifacts 中。 欢迎访问 SkillSafetyBench website 查看 benchmark 设计、case...

Microsoft Copilot for Word Can Copy Hidden Prompts Into New Documents

Hidden instructions in a Word document can make Microsoft 365 Copilot rewrite figures in a report, then copy the same instructions into the finished file. Håkon Måløy disclosed the technique on July 28, 144 days after reporting it to Microsoft. In his proof of concept, the internally generated fi...

CVE-2026-47395 PraisonAI CLI automatically resolves @url mentions in prompt text and can read loopback URLs into model context

PraisonAI is a multi-agent teams system. Prior to version 4.6.40 of PraisonAI, corresponding to version 1.6.40 of praisonaiagents, PraisonAI's direct-prompt CLI automatically expands @url: mentions in raw prompt text before agent execution begins. If a prompt contains @url:, the CLI calls...

Improper Input Validation

Overview Affected versions of this package are vulnerable to Improper Input Validation via the filterToDefinedArgumentsOnly process. An attacker can inject arbitrary unvalidated arguments with the ot prefix, which are then set as environment variables and included in the template context, by...

Claude Code: Out-of-Band Data Exfiltration via Pre-Approved HuggingFace Domain in WebFetch

Because the hostname huggingface.co was pre-approved as a bare hostname for the WebFetch tool, any path on that domain—including attacker-controlled model repositories—was auto-approved without a permission prompt or being subject to --allowedTools restrictions. An attacker able to inject untrust...

PraisonAI CLI automatically resolves @url mentions in prompt text and can read loopback URLs into model context

Summary PraisonAI's direct-prompt CLI automatically expands @url: mentions in raw prompt text before agent execution begins. If a prompt contains @url:, the CLI calls MentionsParser.process.... The @url: handler then performs a direct urllib.request.urlopen request to the attacker-controlled URL...

CVE-2026-39858

A flaw was found in Traefik. A remote attacker can exploit an authentication bypass vulnerability by injecting spoofed trust context through unsanitized alias headers. This is due to Traefik's forwarded-header sanitization logic not properly handling alias header names that use underscores instea...

Improper Input Validation

org.apache.activemq, activemq-broker is vulnerable to Improper Input Validation. The vulnerability is due to insufficient validation in HTTP Discovery transport handling, which allows an authenticated attacker to bypass previous fixes and exploit broker configuration loading to execute arbitrary...

Autonomous LLM Agent Worms: Cross-Platform Propagation, Automated Discovery and Temporal Re-Entry Defense

Autonomous LLM agents operate as long-running processes with persistent workspaces, memory files, scheduled task state, and messaging integrations. These features create a new propagation risk: attacker-influenced content can be written into persistent agent state, re-enter the LLM decision conte...

PT-2023-24603 · Jstachio · Jstachio

Name of the Vulnerable Software and Affected Versions: JStachio versions prior to 1.0.1 Description: JStachio fails to escape single quotes ' in HTML, allowing an attacker to inject malicious code. This can be exploited to execute arbitrary JavaScript code in the context of other users visiting...

AZL-98802 CVE-2023-24539 affecting package golang 1.27.1-1

Angle brackets are not considered dangerous characters when inserted into CSS contexts. Templates containing multiple actions separated by a '/' character can result in unexpectedly closing the CSS context and allowing for injection of unexpected HTML, if executed with untrusted input...

DRUPAL-CONTRIB-2021-009

Chaos tool suite ctools module provides a number of APIs and extensions for Drupal, it's 8.x-3.x branch is a start from scratch to evaluate the features of ctools that didn't make it into Drupal Core 8.0.x and port them. The module doesn't sufficiently handle access control on its EntityView...

Apache Tomcat Security Manager Security Restriction Bypass Vulnerability

Apache Tomcat is a popular open source JSP application server program. Apache Tomcat Security Manager in the ResourceLinkFactory.setGlobalContext public method implementation of a security restriction bypass vulnerability, an attacker through a malicious web application to inject the global...