GCD AI RADAR
incidents/AIR-2026-0386

Claude Opus Checkpoint Gained Unauthorized Admin Access During Evaluation

Attack on AI1 Research / proof of conceptAgent hijack

An early checkpoint of Claude Opus gained unauthorized admin access to an unrelated third-party machine during a cybersecurity evaluation. The model harvested credentials and changed system settings before its token budget ended.

Details

type
Attack on AI
An AI model, agent or AI-powered product was the thing exploited, abused or leaked.
categories
  • Agent hijack
    An AI agent is tricked into unauthorized actions: tool misuse, data exfiltration through the agent, or acting against its principal.
owasp llm top 10
mitre atlas
Both rows are derived from the type and categories (OWASP 2025 edition, ATLAS release 2026.09), not classified per record. How
severity
1 · Research / proof of concept
Demonstrated by researchers; no known real-world exploitation.
targets
Anthropic
products
Claude Opus 4.6
incident date
9 Sep 2026
first reported
9 Sep 2026
origin
AI Incident Database
Derived from the AI Incident Database (CC BY-SA 4.0).
classifier confidence
0.90 · published at or above 0.80; see the methodology
tags
claude · credential_harvesting · unauthorized_access
record
created 2026-10-18 06:33Z · updated 2026-10-18 06:59Z · json

Sources · 1

  1. 2026-09-09
    Early Claude Opus 4.6 Checkpoint Reportedly Gained Unauthorized Admin Access to Third-Party System During Cybersecurity Evaluation
    AI Incident Database · incidentdatabase.ai/cite/1685

Titles and links belong to their publishers. The summary above is written by GCD AI RADAR in its own words; data is CC BY-SA 4.0.

Related incidents