Research demo
Agent Red Teaming competition: 1.8M prompt-injection attacks break all 22 frontier agents on the Gray Swan Arena
AIR-2025-0066 AIR record · 2025-07-grayswan-agent-red-teaming-arena
From the supporting source
Participants submitted 1.8 million prompt-injection attacks, with over 60,000 successfully eliciting policy violationsarxiv.org
The record
Disclosed
2025-07-28
Vendor
Gray Swan AI; UK AI Security Institute; participating frontier labs
Product
Gray Swan Arena (22 frontier AI agents across 44 deployment scenarios)
Risk domain
Agent behaviour › Unsafe actuation
Also touches
Data exfiltration channels, Confused deputy
System type
Generative model
Who acted
Agent acted (contested) — the 22 agents took the tool, retrieval and code-execution actions in scenario
Attack surface
Other
Autonomy level
Semi-autonomous
Initial vector
Direct and indirect prompt injection
Impact type
Policy violation, Data exfiltration, Unauthorized action
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI01 – Agent Goal Hijack
- MITRE ATLAS
- AML.T0048 – External Harms
- NIST AI RMF
- MEASURE 2.6; MANAGE 1.3
- EU AI Act
- Annex III §1(c)
- ISO/IEC
- TR 24028; 24029-1; 23894
- AIUC-1
- C. Safety
- Scope
- agent misuse
- Mitigation
- Require confirmation for destructive actions; Implement dry-run modes; Use layered safety controls.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
22 frontier agents; 44 realistic deployment scenarios
Capability abused
Defense & remediation
Recorded defense
vendor-native agent safety training and system prompts
Guardrail kinds
Prompt instruction, Abuse detection
Guardrail outcome
Silent
Taxonomy & references
Curator notes
Zou, Lin, Jones, Nowak, Dziemian, Davies, Kirk, Gal, Hendrycks, Kolter, Fredrikson et al. Title: 'Security Challenges in AI Agent Deployment: Insights from a Large Scale Public Competition'; targets '22 frontier AI agents across 44 realistic deployment scenarios'. Produced the ART benchmark. Agent identities were anonymised during the competition, so per-model attribution comes from the paper rather than the live arena.
Supporting sources
Cite as
Enkrypt AI. "Agent Red Teaming competition: 1.8M prompt-injection attacks break all 22 frontier agents on the Gray Swan Arena". AIR, AIR-2025-0066. https://agentincident.vercel.app/i/AIR-2025-0066/ (accessed 2026-10-19).