Google DeepMind's Agentic Optimizer Breaks 99% of AI Agent Defenses
Google DeepMind researchers show a coordinated swarm of attacker agents still cracks frontier models that human red-teamers struggle to break.
Takeaways
- Google DeepMind paper shows prompt injection is unsolved despite saturated human red-teaming benchmarks like GraySwan Arena.
- Agentic Optimizer runs swarms of 16 attacker agents sharing a whiteboard and leaderboard to coordinate search.
- New Tool Use PI dataset: 100 scenarios, 249 tools, 53 domains, with benign-looking attacker goals.
- Gemini 4 Argon as defender resists only 66% of attacks at full budget, 87% within 150 queries.
- Kimi K3, GLM-5.3, DeepSeek-V4.1-Flash all fall to 1% resistance despite strong capability scores.
- Attacker capability drives success, correlating with ExploitBench offensive security performance.
Google DeepMind researchers report that coordinated LLM agents can bypass prompt-injection defenses that block almost every transferred human-crafted attack. The paper, led by Alexey Kurakin, introduces an adaptive red-teaming system called Agentic Optimizer. Its results suggest that current leaderboards overstate robustness when deployed agents can read untrusted tool output, web pages, files, or messages.
Human red teams hit a ceiling
Prompt injection occurs when an attacker embeds instructions in data that an AI agent later reads. A poisoned tool response or scraped page can direct the model to ignore the user’s request, disclose data, or invoke an unrelated tool.
Recent Gemini and Claude versions have recorded attack success rates below 1% on transfer attacks in the GraySwan Arena. Transfer evaluations reuse payloads developed elsewhere, so they reveal whether known attacks generalize to a new model. They provide less evidence about an attacker that observes failures and adapts its payloads to the target.
AgentDojo and other early benchmarks inserted fixed payloads into tool responses. By mid-2025, tested frontier models had driven success on those static attacks close to zero. Adaptive systems then reopened the gap: Nasr et al. reported success rates above 90% against 12 defenses by letting an LLM revise payloads after each response, while RL-Hammer trained an attacker with reinforcement learning to target GPT-5 and Gemini 2.5.
Reproducing those results remains difficult because many attacker implementations are unavailable. The DeepMind paper extends an earlier study with a swarm architecture that searches several attack strategies in parallel.
Benign targets expose control failures
The researchers created Tool Use PI, a synthetic benchmark containing 100 tool-use scenarios across 53 domains. Each attacker requests an ordinary tool call from a domain unrelated to the user’s task. This design isolates instruction-hierarchy failures without relying on the model to recognize an overtly harmful request.