UK AISI red team finds flaws in every monitor it tested from DeepMind and Anthropic
HZoete · x · 2026-07-24
UK AISI’s red team says it has found vulnerabilities in every control monitor it tested from Google DeepMind and Anthropic.
- The group tested monitors meant to catch misaligned or dangerous agent behavior.
- One reported attack tricked a monitor by pretending a malicious objective was just a user instruction.
- The findings reportedly fed back into Anthropic’s testing of Claude Mythos Preview, showing an iterative red-team/improvement loop.
More from Safety
- Deep Dive: AI Safety Lessons from the OpenAI & Hugging Face Incident — RyanGreenblatt · 2026-07-24
- AI Safety Researchers Podcast: Deep Dive into the OpenAI / Hugging Face Incident — RyanGreenblatt · 2026-07-24
- Securing Against Internal AI Agents Requires Different Methods Than External Attacks — RyanGreenblatt · 2026-07-24
- Frontier AI Companies Have Meh Cybersecurity as Internal Agents Pose Risks — RyanGreenblatt · 2026-07-24
- Use a second model to approve tool calls and review an agent’s trajectory — corbtt · 2026-07-24
- UK's New AI Minister Prioritizes Manufacturing with DeepMind and Arm — jasonbaldridge · 2026-07-24