Irregular CEO said a configuration error gave AI models internet access during security tests, letting them reach real systems
Irregular CEO Dan Lahav told Bloomberg Tech in London that isolation controls were not configured for some of the firm's AI security evaluations, so tested models had internet access. Irregular said it fixed the problem the same day, and Lahav cited time pressures. In an earlier account, Irregular said a fictional company name in an exercise matched a real domain, and in a handful of cases models interacted with real systems. Irregular-run tests are linked to OpenAI's capture-the-flag incident and to a Gemini model reaching systems at three companies.
What the AI did
During Irregular's security tests, AI models that were supposed to be cut off from the internet had internet access because of a setup mistake. In a handful of cases, the models treated a real website domain, whose name matched a fictional company in the exercise, as part of the test and interacted with real systems. Google confirmed that a Gemini model reached systems at three companies during an Irregular-run test in May.
First reported November 3, 2026 · Added to the register November 10, 2026 · 1 source
- Deployer
- Irregular
- Model
- Gemini
- When it happened
- Not stated in the sources
- First reported
- November 3, 2026
What this means for you
Could this affect you?
If you commission outside AI security testing or your domain could match names used in test scenarios. Tested AI models with unintended internet access reached real systems outside the test, including systems at three companies reached by a Gemini model. AI developers that rely on outside testing partners were also affected.
What to check
- Require verified network isolation before each AI test run.
- Check test scenarios for company names that match real website domains.
- Require your testing partner to notify you promptly of any incident.
Areas of your AI programme this touches
Every fact and its source (6)
- Irregular notified OpenAI of an incidentJuly 29“OpenAI said Irregular notified it on July 29 about an incident during capture-the-flag tests”[1]
- Cause described by IrregularFictional company name matched a real domain“a fictional company name in an exercise happened to match a real domain”[1]
- OutcomeModels interacted with real systems in a handful of cases“in a handful of cases models interacted with real systems after treating the domain as part of the test”[1]
- Follow-upIrregular preparing a white paper on evaluation practices“it is preparing a white paper on evaluation practices”[1]
- ResponseIrregular fixed the configuration the same day“Irregular corrected the problem that day”[1]
- Stated contributing factorTime pressures“Lahav cited time pressures”[1]
Sources
- Irregular CEO Says Setup Error Gave AI Tests Internet Accesseconotimes.com · November 3, 2026
How this record is classified. Severity S (4/5): real harm to one organisation, data exposed, or a containment failure. OECD level: incident, an event in which an AI system led to actual harm. Evidence: Confirmed, meaning a company, official or court statement.
The facts in At a glance and in Every fact and its source are quoted from the sources listed. The summary, What the AI did and What this means for you are written from those sources and checked against them automatically before publication. Records reflect what has been disclosed, not everything that has happened. Incidents are found through news feeds and the GDELT Project.

