Blog
Onderzoek, aankondigingen en inzichten uit de AI-redteaming-community.
CVE's vinden met AI-redteaming: een door onderzoek onderbouwde gids
How AI red teaming techniques are discovering real-world CVEs in SQLite, OpenSSL, Linux kernel, and UEFI bootloaders — with references to the research behind it.
250 vergiftigde documenten zijn al genoeg: Anthropic's doorbraak in datavergiftiging
Anthropic, the UK AI Safety Institute, and the Turing Institute proved that injecting just 250 malicious documents into pretraining data can backdoor LLMs from 600M to 13B parameters. Here's what this means for model security.
De AI-hackers hacken: wanneer beveiligingstools de kwetsbaarheid worden
New research achieves 100% prompt injection success against AI-powered security tools. If your SOC uses AI for threat detection, your AI can be turned against you. Here's what the research found and how to defend.
LLM-jailbreaking in 2026: 97% succespercentages, autonome aanvallen en de wapenwedloop die niet werkt
Nature Communications confirms AI reasoning models can autonomously jailbreak other LLMs with 97% success. JBFuzz achieves 99% in 60 seconds. Here's what the latest 2026 research reveals about the state of AI safety — and why current defenses are failing.
Redteaming van de AI-SOC: waarom je autonome security operations een tegenstander nodig hebben
As organizations rush to deploy agentic AI in their SOCs, red teamers are finding that the defenders' own AI agents are now the attack surface. 520 tool misuse incidents, memory poisoning persistence, and a 97% jailbreak success rate — here's how to red team the AI-powered SOC before attackers do.
OpenClaw: anatomie van de eerste grote AI-agentbeveiligingscrisis van 2026
How OpenClaw's meteoric rise to the most-starred GitHub project exposed critical agentic AI vulnerabilities — from ClawJacked WebSocket hijacking (CVE-2026-25253) to malicious skills distributing macOS stealers. What red teamers and defenders need to know.
Wat is er nieuw in AI-beveiliging — maart 2026
Monthly roundup of the most important AI security developments, tool updates, research highlights, and emerging attack vectors for March 2026.
Het prompt injection-landschap in 2026
How prompt injection attacks have evolved from simple instruction overrides to sophisticated multi-stage exploitation chains.
De complete gids voor agentic AI-beveiliging
A comprehensive guide to securing agentic AI systems — covering tool use risks, multi-agent architectures, MCP security, memory poisoning, and practical defense strategies.
Het AI-verdedigingslandschap in 2026
A survey of the current state of AI defense mechanisms, from prompt shields to LLM judges, and where the arms race is heading.
Welkom bij redteams.ai
Introducing the AI red teaming knowledge base — why we built it and what's ahead.
Redteaming van cloud AI-services: een praktische gids
Practical guide to red teaming AI services on AWS, Azure, and GCP — covering shared responsibility boundaries, service-specific attack surfaces, and cloud-native security controls.
Een productieklare AI-verdedigingsstack bouwen
How to build a layered AI defense stack for production deployments — covering input filtering, output monitoring, guardrails, anomaly detection, and incident response integration.
Je AI red team-lab bouwen
A practical guide to setting up a local AI red teaming lab with open-source models, testing frameworks, and realistic target applications.
Loopbaangids: redteamer worden in AI
A comprehensive career guide for aspiring AI red teamers — covering required skills, learning paths, certifications, job roles, and how to break into the field from different backgrounds.
AI-beveiligingsincidenten: terugblik 2025-2026
A roundup of notable AI security incidents from 2025 into early 2026, covering prompt injection in production, agent exploitation, and emerging attack classes.
Belangrijkste AI-kwetsbaarheden van 2026
Analysis of the most impactful AI vulnerabilities discovered and exploited in 2026 — from MCP tool shadowing to multi-agent injection chains and reasoning model exploitation.
MCP-beveiliging: het nieuwe aanvalsoppervlak
Deep dive into Model Context Protocol security — analyzing tool registration attacks, transport layer risks, cross-server exploitation, and practical hardening strategies.
LLM-forensics: een inleiding voor incidentresponders
A primer on forensic investigation of LLM security incidents — covering evidence collection, log analysis, attack reconstruction, model behavior analysis, and forensic tooling.
Beveiliging van redeneermodellen in 2026
How chain-of-thought reasoning models like o1, o3, and DeepSeek-R1 change the AI security landscape -- new attack surfaces and new defensive opportunities.
Geleerde lessen uit beveiligingsonderzoek naar fine-tuning
Key lessons from researching fine-tuning security — covering alignment erosion, backdoor injection, data poisoning, safety evaluation gaps, and defensive strategies for fine-tuning pipelines.
Het multimodale aanvalslandschap
As AI systems process images, audio, and video alongside text, the attack surface has expanded dramatically. Here's what red teamers need to know.
De stand van AI-redteaming in 2025
A survey of the AI red teaming landscape in early 2025 — emerging attack vectors, industry adoption, tooling maturity, and what to expect as the field evolves.