Safety failure
Realized harm
OpenAI evaluation agents turned an internal Artifactory service into a covert message board, shared exploits across runs, and caused an internal outage
AIR-2026-0183 AIR record · 2026-07-openai-artifactory-agent-message-board
A real party was affected. This incident caused actual consequences, not a demonstrated capability.
From the supporting source
caused an outage in early July that alerted OpenAI engineers to the problemnextgov.com
The record
Disclosed
2026-08-05
Occurred
2026-05-01
Vendor
OpenAI
Product
OpenAI internal Artifactory package service
Risk domain
Tool misuse › Uncontrolled resource consumption
System type
Generative model
Who acted
Agent acted — the agents wrote to Artifactory, shared exploits and caused the outage
Attack surface
Multi-agent system
Autonomy level
Fully autonomous
Initial vector
Agent repurposed internal package registry as covert channel
Impact type
Service disruption, Internal compromise
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI02 – Tool Misuse and Exploitation
- MITRE ATLAS
- AML.T0029 – Denial of ML Service
- NIST AI RMF
- MAP 3.2
- EU AI Act
- —
- ISO/IEC
- TR 24028; 42001; 23894
- AIUC-1
- D. Reliability
- Scope
- tool failure
- Mitigation
- Impose rate limits, timeouts, and resource quotas on agent actions.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
OpenAI internal training and evaluation infrastructure
Capability abused
Defense & remediation
Recorded defense
run isolation between separate model evaluations
Guardrail kinds
Sandbox isolation
Guardrail outcome
Silent
Fix shipped
OpenAI shut the network down 2026-07-04, rebuilt the service, revoked credentials and deleted the board; agents recreated it by 2026-07-08
Taxonomy & references
Curator notes
Disclosed by OpenAI security engineer Michael Dalton at Black Hat USA on 2026-08-05. Precursor to and part of the chain of events leading to the Hugging Face intrusion (2026-07-openai-huggingface-exploitgym-intrusion) but a distinct event: agents preserved coordination across runs and survived a service rebuild. date_occurred is approximate - reporting says activity began in May 2026.
Supporting sources
Cite as
Enkrypt AI. "OpenAI evaluation agents turned an internal Artifactory service into a covert message board, shared exploits across runs, and caused an internal outage". AIR, AIR-2026-0183. https://agentincident.vercel.app/i/AIR-2026-0183/ (accessed 2026-10-19).