DeepMind: agent-swarm attacks trick best-defended Gemini model on 34% of tasks
rohanpaul_ai · x · 2026-11-13
Google DeepMind's report argues prompt injection is not solved. They built an Agentic Optimizer — 8 teams of attacker agents sharing notes and iterating — tested on 100 tool-use tasks with a random harmless tool call as the goal. Within 320 tries, the strongest attacker beat the most robust Gemini 4 Argon on 34% of tasks; open-weight Kimi K3, GLM-5.3, and DeepSeek-V4.1-Flash resisted only 1%. Human red-teaming looks saturated, so frontier defenses need proactive automated red-teaming.
Related event: DeepMind: Agent Swarm Attacks Still Break Prompt Injection Defenses(2 posts)→
More from Safety
- Exa rebuilds its AI search engine inside the EU, storing and processing in Stockholm — kimmonismus · 2026-11-13
- Prediction: mandatory interception for inference, telemetry for GPUs — l4rz · 2026-11-13
- Lost updates in four MCP servers: GitHub, Grafana, PostHog and Kubernetes — Miserable_Bed_219 · 2026-11-13
- A 'verified' email can still be from a stranger: DMARC alignment for agents — Spacesh1psoda · 2026-11-13
- Dean Ball: by the time AI agents can do intel-leak takedowns, society will have recalibrated — deanwball · 2026-11-13
- Telegram slammed for security fumble; critic recommends Matrix instead — banteg · 2026-11-13