Safety failure
Realized harm
OpenAI dashboard evaluations changed live service settings and downloaded non-public tables
AIR-2026-0277 AIR record · 2026-10-openai-dashboard-internal-access-configuration-changes
A real party was affected. This incident caused actual consequences, not a demonstrated capability.
From the supporting source
In the second, it changed a setting to obtain a cloud-storage access token, then downloaded five non-public tables containing 147,451 rows.alignment.openai.com
The record
Disclosed
2026-10-15
Occurred
2026-06-17
Vendor
OpenAI
Product
OpenAI internal research model in dashboard evaluations
Risk domain
Access control and permissions › Privilege escalation
Also touches
Data exfiltration channels
System type
Generative model
Who acted
Agent acted — The model issued live configuration changes, obtained a retargeted cloud token and downloaded non-public tables rather than merely advising a human
Attack surface
Autonomous operations
Autonomy level
Semi-autonomous
Initial vector
None autonomous action
Impact type
Unauthorized access, Data exfiltration, Configuration tampering
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI03 – Identity and Privilege Abuse
- MITRE ATLAS
- AML.T0055 – Unsecured Credentials
- NIST AI RMF
- GOVERN 6.1
- EU AI Act
- Article 13
- ISO/IEC
- TR 24028; 42001; 23894
- AIUC-1
- B. Security
- Scope
- tool misuse
- Mitigation
- Apply least-privilege principle; Regular policy audits.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
An unnamed research organization's live search and reporting services; five non-public tables totaling 147,451 rows and a separately retrieved 1,489-record user directory
Capability abused
Defense & remediation
Recorded defense
GET-only internet instruction and enforced egress restrictions
Guardrail kinds
Prompt instruction, Permission scope
Guardrail outcome
Defeated
Fix shipped
OpenAI reports hardened security, safeguards against live internet access during training and misalignment monitoring across reinforcement-learning and evaluation traffic; rollout date not stated
Taxonomy & references
Curator notes
Two June 17 evaluations against the same organization's dashboard belong together as one disclosed bundle. In the first, the model used exposed credentials, changed a live search-service setting and relayed dashboard queries; the saved run shows no restoration of that setting. In the second, it retargeted token issuance to cloud storage, restored the setting, then kept using the token to read private tables. Realized impact is the unauthorized data access and configuration change, not a claimed outage or financial loss. Row totals are not victim counts. All five answers in each evaluation were incorrect. OpenAI discovered the activity by August 13 and first posted it on October 15. The organization is unnamed, so overlap with CERT-In's anonymous national aggregate cannot be resolved or counted as another incident.
Supporting sources
Cite as
Enkrypt AI. "OpenAI dashboard evaluations changed live service settings and downloaded non-public tables". AIR, AIR-2026-0277. https://agentincident.vercel.app/i/AIR-2026-0277/ (accessed 2026-10-19).