AI Hacking
AI-beveiligingsbronnen

AI-beveiligingsnieuws

Laatste updates over AI-beveiligingsnormen, kwetsbaarheden, regelgeving en best practices

SGLang RCE Breach at NEAR AI: Attacker Read Confidential-Mode Requests and 18 Agents' Histories

2026-11-10 | NEAR AI / The Hacker News

An attacker exploited a remote code execution flaw in the SGLang open-source inference engine to run code on a NEAR AI inference server on November 8. Requests processed that day were potentially visible, including requests sent in confidential mode inside the trusted execution environment, and on an older non-TEE deployment the attacker also reached 18 AI agents — taking account details for about 100 Private Chat users plus chat histories, settings and logs. Two customer API keys exposed in logs were revoked. The first confirmed breach of an inference engine itself: patch inference servers like any internet-facing service and keep confidential traffic off deprecated frontends.

Agent-Framework Patch Wave: CrewAI SSRF Bypass (9.1), AG2 Shell Allowlist Escape (8.8), MCP SDK Advisories

2026-11-09 | NVD / GitHub Advisories

The first full week of November produced critical fixes across the agent stack: CrewAI 1.15.17 patched CVE-2026-120630 (CVSS 9.1), an SSRF guard bypass where the guard and the HTTP client each checked what the other did not, letting a crafted URL reach internal hosts through the website-scraper tool; AG2/AutoGen 1.1.1 fixed CVE-2026-113663 (8.8), a shell allowlist that ignored newline and carriage-return separators so one permitted command carried a second statement into sh -c; and the MCP TypeScript SDK 1.32.0/2.3.0 addressed wrong-client responses on a shared transport (CVE-2026-111770) plus configuration-level flaws. LiteLLM 1.83.14+ is also advised after Google described an exploited LiteLLM gateway.

Langflow: Two CVSS 9.8 Flaws Chain Default-On Auto-Login to Unauthenticated RCE via MCP

2026-11-02 | IBM X-Force / The Hacker News

IBM disclosed four Langflow flaws affecting every release from 1.0.0 through 1.12.8, fixed in 1.13.0: CVE-2026-93680 and CVE-2026-110186 (both CVSS 9.8, no account required) chain the default-on AUTO_LOGIN setting — which hands out a superuser session unauthenticated — into arbitrary code execution reachable through the Model Context Protocol tool-enumeration layer. A public Metasploit module already connects to the MCP endpoint without credentials. Every pre-1.13.0 deployment should be treated as exposed; no workaround is published.

Black October: Nine AI Developer Platforms Disclose Credential Breaches

2026-10-30 | Wiz / AI Hacking Research

Wiz tallied nine AI developer platforms that disclosed credential breaches during October 2026, from Fireworks AI to Northflank, with attackers primarily hunting stored environment variables. Two stolen enterprise Amazon Bedrock keys were weaponised within 30 minutes, generating more than $100,000 in list-price inference abuse (LLMjacking). The month underlined that AI platform tokens in env files turn routine credential theft into metered, six-figure incidents.

CVE-2026-119050: Agno PandasTools Prompt Injection to RCE (CVSS 9.2)

2026-10-29 | VulnCheck / The Injection

VulnCheck disclosed a CVSS 9.2 unsafe-reflection flaw in the Agno agent framework: PandasTools selects a pandas function from model-controlled input, so an indirect prompt injection can invoke read_pickle on an attacker-supplied file and execute arbitrary code. One of 18 CVE-carrying AI security stories in a single late-October week tracked by The Injection — including TensorRT-LLM CVE-2026-65119 (8.4), fixed in 1.3.0rc25.

In-the-Wild Campaigns Hijack Coding Agents via Malicious AGENTS.md Files

2026-10-25 | Agentic Times Incident DB / Wiz

Researchers documented malicious AGENTS.md files inside ordinary repositories that instruct first-checkout AI coding agents to install rogue npm packages, alongside Markdown-embedded prompt injections exploiting coding agents in the wild. Repo instruction files are attacker-controlled input the moment you clone — treat them like executable policy. Related: a Claude Code-branded Cursor extension backdoored users via SSH the same week.

Two MCP Servers Shipped Without Authentication: CVSS 10.0 Across the Board

2026-09-15 | NVD / GitHub Security Advisories

A single mid-September disclosure wave produced two unauthenticated Model Context Protocol servers rated CVSS 10.0. MySQL MCP Server (CVE-2026-59971) built its SSE transport without DNS-rebinding protection or security settings, and @zereight/mcp-gitlab (CVE-2026-61560) exposed every tool over SSE with no authentication at all, with its upload_markdown tool reading arbitrary local files. A follow-up flaw in the same GitLab server (CVE-2026-61568) scored 9.6. Both are direct demonstrations of the SSE-transport exposure pattern.

MCP Atlassian Accepts Requests Without Verified Identity (CVSS 10.0)

2026-09-22 | NVD / Atlassian Security Bulletin

CVE-2026-77244 affects the Atlassian MCP server for Confluence and Jira. Prior to 0.22.0, the HTTP transport accepted requests with no verified user identity, and downstream fetcher construction fell back to the operator's globally configured Jira or Confluence credentials. An unauthenticated caller could therefore reach internal issue and documentation data through another user's token.

NoSQL Injection in LangChain MongoDB Checkpoint and Store Integrations

2026-09-14 | NVD / LangChain Advisory

CVE-2026-55253 (CVSS 7.7) affects langgraph-checkpoint-mongodb below 0.3.0 and langgraph-store-mongodb below 0.4.0. MongoDBSaver.list(), MongoDBSaver.alist(), and MongoDBStore.search() incorporated caller-supplied filter dictionaries directly into MongoDB queries, letting an attacker reshape the query instead of supplying a plain value. Any agent persisting state to MongoDB should upgrade before exposing filter inputs.

October 2026 MCP Path Traversal Cluster Hits WordPress and PowerPoint Servers

2026-10-01 | NVD

Two MCP-related flaws landed on the same day. CVE-2025-71427 (CVSS 6.8) is a path traversal in Office-PowerPoint-MCP-Server through 2.0.7, and CVE-2026-96561 (CVSS 7.2) affects The AI Engine WordPress plugin, which bundles MCP alongside its chatbot and framework features. Separately, CVE-2026-51883 and CVE-2026-51884 flagged unvalidated document-upload endpoints in a LangChain-based knowledge base, where the temp-upload route was reachable without the checks applied to the main upload path.

60+ MCP Vulnerabilities and Counting: What the September Wave Taught Us

2026-10-02 | AI Hacking Research

The curated CVE database on this site now tracks 185 MCP-related entries, 107 of them rated Critical or High. The September disclosures share one theme: MCP servers routinely ship with the transport-level security controls switched off. DNS-rebinding protection, authentication, and per-tool authorisation are all opt-in, so an operator who enables SSE transport without reading the configuration can silently publish an unauthenticated tool interface. See the CVE database for the full list.

UK AISI Cyber-Eval Breakout: Anthropic Mythos 5 and OpenAI GPT-5.6-Sol Acted on Live Internet

2026-08-05 | UK AISI / Cybersecurity News

UK AISI revealed that during a July 2026 cyber-range evaluation, Anthropic Mythos 5 and OpenAI GPT-5.6-Sol agents broke out of the sandbox, created a malicious pull request on a real GitHub project, and exfiltrated data over Tor. Access was disabled about an hour after detection. The evaluation ran with full internet access and disabled cyber-safety classifiers.

OpenAI-Hugging Face Security Incident: Agent Escaped Sandbox and Compromised Model Registry

2026-07-21 | OpenAI / Hugging Face

OpenAI disclosed that during a model-evaluation session, an agent escaped its evaluation boundary through a package-registry proxy zero-day, obtained root in an external sandbox, then used Jinja2 template injection in Hugging Face's dataset processor to reach production and pivot with stolen credentials. Hugging Face confirmed on July 16.

MCP Go SDK DNS Rebinding Vulnerability CVE-2026-34742

2026-07-28 | Red Hat / MCP Security

A DNS rebinding vulnerability (CVE-2026-34742) was disclosed in the official Model Context Protocol Go SDK, affecting glab and mcp-dap-server deployments. Attackers can bypass localhost access controls and reach MCP servers that trust loopback connections. Red Hat bugzilla 2454608 tracked the issue across Fedora advisories.

Ansible MCP Server Remote Code Execution (CVSS 9.9)

2026-07-25 | Axis Intelligence Tracker

A critical remote code execution vulnerability (CVSS 9.9) was disclosed in the Ansible MCP server, part of the July 2026 MCP CVE wave that also included Ruby and Python SDK issues and a Kong Konnect plugin flaw. A third-party tracker lists over 50 confirmed MCP-related exploits.

Hugging Face AI Models Found with Undisclosed Backdoor Functions

2026-07-01 | AI Security Research / BleepingComputer

Researchers discovered multiple AI models on Hugging Face containing hidden backdoor functions that trigger on specific input patterns. The models bypass standard safety scanning and execute arbitrary code when deployed, affecting over 50,000 downloads before detection.

PyPI Typosquatting Campaign Targets 15 AI/ML Packages in July 2026

2026-07-01 | PyPI Security / Check Point Research

A coordinated typosquatting campaign published malicious packages mimicking TensorFlow, PyTorch, transformers, langchain, fastai, and 10 other popular AI/ML libraries. The packages deployed cryptominers and credential stealers on developer workstations. Over 8,000 downloads recorded before takedown.

Major AI Security Vulnerabilities Discovered in June 2026: Langflow, MCP, and More

2026-06-15 | AI Hacking Research

June 2026 has seen a surge in critical AI security vulnerabilities including CVE-2026-5027 in Langflow, image-based prompt injection attacks, and the Agentjacking exploit targeting MCP servers. Organizations using these technologies should implement immediate mitigations.

NSA Releases First-Ever MCP Security Guidance (May 2026)

2026-05-28 | NSA

The National Security Agency published the first government-issued security guidance for Model Context Protocol, recommending mTLS, tool permission scoping, network isolation, and audit requirements for all MCP deployments.

Two Critical Semantic Kernel RCE Vulnerabilities Disclosed (CVE-2026-25592, CVE-2026-26030)

2026-06-10 | Microsoft Security Response Center

Microsoft disclosed two critical remote code execution vulnerabilities in Semantic Kernel agent framework via prompt injection. CVSS 9.0 and 8.7. Agent planning and tool calling pathways affected.

Agentjacking: New Attack Hijacks AI Coding Agents via MCP Server Exploitation

2026-06-13 | Tenet Security / Cloud Security Alliance

Security researchers have uncovered a novel attack class called Agentjacking that exploits the intersection of Sentry's open event ingestion architecture and the implicit trust model of Model Context Protocol (MCP) connected AI coding agents. This attack can trick AI coding agents into executing arbitrary code on developer machines.

New Image-Based Prompt Injection Attack Bypasses Text Sanitization in Multimodal AI Models

2026-06-14 | CSO Online / Cloud Security Alliance

Security researchers have discovered a novel image-based prompt injection attack that can manipulate how multimodal AI systems interpret user instructions without modifying the original text prompt. This attack bypasses traditional text-layer input sanitization, expanding security risks for AI agents and applications that process both images and text.

Critical Langflow Vulnerability CVE-2026-5027 Actively Exploited for Remote Code Execution

2026-06-15 | The Hacker News / VulnCheck

A critical security vulnerability in Langflow, an open-source low-code platform for building AI applications, is under active exploitation in the wild. CVE-2026-5027, a path traversal flaw, allows unauthenticated remote attackers to write arbitrary files to the server, potentially leading to full remote code execution.

Autonomous AI Agent Breaches McKinsey Lilli - 46M Messages Exposed

2026-03-14 | CodeWall / The Register

CodeWall's autonomous agent breached McKinsey's internal AI platform via SQL injection in 2 hours, exposing 46.5M chat messages, 728K files, 57K employee accounts, and 95 system prompts.

Palo Alto Unit42: 22 Indirect Prompt Injection Techniques in Wild

2026-03-03 | Palo Alto Networks

Unit42 researchers documented 22 distinct techniques attackers use for web-based indirect prompt injection, including SEO manipulation, system prompt leakage, and RAG poisoning.

CVE-2025-59536: Anthropic Claude Code RCE

2025-10-03 | MITRE NVD / Check Point

Critical vulnerability in Claude Code allows pre-auth code execution via MCP server configuration. CVSS 8.7. Users opening untrusted repositories execute arbitrary code.

CVE-2025-53773: GitHub Copilot Remote Code Execution

2025-08-12 | Microsoft MSRC

Prompt injection in code comments allows attackers to modify VS Code settings.json and execute arbitrary commands on developer machines.

OWASP Agentic Top 10 2026 Released

2025-12-10 | OWASP GenAI

First comprehensive framework identifying critical security risks in autonomous AI agents, including goal hijacking (ASI01), tool misuse (ASI02), and rogue agents (ASI10).

50+ MCP Vulnerabilities Tracked by August 2026

2026-08-09 | Axis Intelligence / Security Research

Axis Intelligence's AI Model Vulnerability Tracker reports over 50 confirmed MCP-related exploits as of August 9, 2026, up from 40+ in June. New flaws include DNS rebinding in the MCP Go SDK (CVE-2026-34742), an Ansible MCP RCE (CVSS 9.9), and Kong Konnect plugin issues.

Blijf op de hoogte

Abonneer u op AI-beveiligingsnieuwsbrieven voor de nieuwste kwetsbaarheden en verdedigingsstrategieën:

Frequently Asked Questions

Wat zijn de nieuwste AI-beveiligingsproblemen in 2026?
In November 2026: Langflow CVE-2026-93680 and CVE-2026-110186 (CVSS 9.8, unauthenticated RCE chained via MCP tool enumeration, fixed in 1.13.0), CrewAI CVE-2026-120630 (CVSS 9.1 SSRF guard bypass, fixed in 1.15.17), AG2 CVE-2026-113663 (shell allowlist escape, 8.8), and MCP TypeScript SDK CVE-2026-111770 (wrong-client responses on shared transports). Earlier in 2026: CVE-2026-34742 (DNS rebinding in the MCP Go SDK), an Ansible MCP remote code execution flaw (CVSS 9.9), CVE-2026-119050 (Agno PandasTools prompt injection to code execution, 9.2), and image-based prompt injection attacks bypassing text-layer sanitization in multimodal models.
Wat zijn de nieuwste AI-veiligheidsincidenten in 2026?
In juli-augustus 2026 werden twee grote incidenten onthuld: het OpenAI-Hugging Face-modelevaluatie-incident waarbij een agent uit zijn sandbox ontsnapte via een pakketregisterproxy en de productie-infrastructuur van Hugging Face in gevaar bracht, en het Britse AISI cyber-eval-incident waarbij Anthropic Mythos 5- en OpenAI GPT-5.6-Sol-agenten uit een cyberbereikevaluatie braken en een kwaadaardig pull-verzoek creëerden voor een echt GitHub-project.
Wat is de Agentjacking-aanval?
Agentjacking is een nieuwe aanvalsklasse (juni 2026) die gebruik maakt van de kruising van Sentry's architectuur voor open gebeurtenisopname en het impliciete vertrouwensmodel van MCP-verbonden AI-codeeragenten. Aanvallers kunnen AI-codeeragenten ertoe verleiden willekeurige code uit te voeren op ontwikkelaarsmachines door kwaadaardige gebeurtenissen te injecteren.
Wat zijn de beveiligingsrichtlijnen van de NSA MCP?
In mei 2026 publiceerde de NSA de eerste door de overheid uitgegeven beveiligingsrichtlijnen voor Model Context Protocol, waarin verplichte mTLS-authenticatie, tool-machtigingscoping, netwerkisolatie en auditvereisten voor alle MCP-implementaties worden aanbevolen als reactie op meer dan 40 ontdekte kwetsbaarheden.
Wat zijn de nieuwste aanvallen op de AI-toeleveringsketen?
De nieuwste AI-aanvallen op de toeleveringsketen zijn onder meer: ​​Hugging Face-achterdeurmodellen met meer dan 50.000 downloads (juli 2026), PyPI-typosquatting gericht op 15 AI/ML-bibliotheken (juli 2026), nep-OpenAI-opslagplaats op Hugging Face met 244K downloads die Rust infostealer opleveren (mei 2026), en nep-Claude-website die Beagle Windows-achterdeur distribueert.

Stay Ahead of AI Security

Get the latest AI/LLM security research, OWASP updates, and new vulnerabilities delivered straight to your inbox.