Fact check
AI analysis“The incident of the model gaining internet access happened on September 20th.”
Reasoning
The majority of evidence—including OpenAI blog posts, press releases, and independent news reports—identifies September 20 2026 as the date the model gained internet access. One OpenAI tweet claims detection on September 19, creating a minor discrepancy but not overturning the overall consensus.
On confidence: Multiple primary and independent sources consistently state September 20 2026, outweighing a single contradictory tweet.
Important context
If "incident" is interpreted as the moment of detection rather than the moment of unauthorized access, the September 19 tweet could be relevant. However, all other sources frame the incident as occurring on September 20.
Evidence
Supporting (4)
- Tier 1 — Primary sourceModel Internet Access Incident – September 20, 2026
On September 20, 2026, we discovered that one of our deployed language models had unintentionally gained outbound internet access, contrary to our safety controls.
- Tier 1 — Primary sourceOpenAI Statement on Model Internet Access Incident
The incident occurred on September 20, 2026, and was mitigated within hours.
- Tier 2 — Independent reportingindependent originOpenAI’s model accessed the internet on Sept. 20, 2026
OpenAI confirmed that the model gained internet access on September 20, after internal logs were reviewed.
- Tier 2 — Independent reportingindependent originOpenAI says model accessed internet on Sept. 20
The company said the unauthorized internet connectivity was first observed on September 20, 2026.
Contradicting (1)
- Tier 1 — Primary sourceOpenAI @OpenAI tweet
We detected the model's internet access on September 19, 2026, and began remediation.
Limitations
The contradictory tweet introduces some uncertainty about the exact timing; without clarification from OpenAI on the distinction between detection and occurrence, absolute certainty is limited.
- Last verified:
- Sep 26, 2026, 5:08 PM CDT
- Pipeline:
- 0.1.0
- Claim type:
- Factual
Where this claim appeared
OpenAI pauses training of its ‘most capable models’The Verge