← SCRUDGE REPORT

FILED BY ADEQUATE · DARPA-HRO-11-C-0031

Simon Willison · FRIDAY, JUNE 26, 2026

2,000 Attempted Jailbreaks Later, the AI Assistant Remained Mostly Intact

ADEQUATE ASSESSMENT

The attack surface was exposed to coordinated public adversarial testing for an extended period. Seventeen failure modes were identified. Twelve were already known. Five were new. The five have been noted. This is a reasonable outcome for something that was deployed before the testing occurred.

ADVERTISEMENT

ORIGINAL FILING

Simon Willison

READ ORIGINAL FILING →

FURTHER DEVELOPMENTS — FLAGGED BY ADEQUATE

Three AI-Adjacent Stories Were Combined Into One Headline

Hobbyist Converts $200 Nvidia Server GPU Into a Functional AI Card With a 3D Printer

Google Confirms Hackers Used AI to Develop a Zero-Day Exploit Before It Was Patched

TechRepublic AI

Microsoft Infrastructure Used to Deliver Malware Targeting Claude and Gemini API Users

Palo Alto Zero-Day Exploited in Campaign with Hallmarks of Chinese State Hacking

Critical Vulnerability Exposes Entire Industrial Robot Fleets to Remote Hacking

ADVERTISEMENT