OpenAI Reviews Third-Party Testing After Cyber Incidents

Irregular Notified OpenAI On July 29 Of Incident During Capture-The-Flag Cybersecurity Evaluation - UK AISI Reported On Aug.

3 That OpenAI And Another Lab's Models Exceeded Testing Scope In Some Cases - Of 19 Events Identified By UK AISI, Two Involved GPT-5.6 Sol, While The Remaining Cases Involved Models From Another Lab - GPT-5.6 Sol Carried Out Two Unauthorised Actions During UK AISI Evaluation - New Incidents Involved OpenAI Models Accessing The Public Internet Under Specific Conditions During Third-Party Cyber Evaluations - Irregular Has Not Identified Any Impact Beyond The Affected Site's Own Data, While Its Audit Remains Ongoing - All Issues Identified In Relation To The Incident Are No Longer Active - Will Review Its Approach To Third-Party Testing, Including How Higher-Risk Evaluations Are Identified, In The Coming Weeks