Cyber access gate keyed to a 1996 US export list

OpenAI’s identity check for Trusted Access for Cyber now rejects government IDs from 44 countries and territories where the company officially sells ChatGPT. That finding comes from George Lubaretsi’s audit, published on August 28. Lubaretsi is a Georgian security researcher and the founder of FOI Security. The refusal happens at the very first screen: choose Georgia in the country selector of Persona, the identity vendor OpenAI uses, and the list of accepted document types comes back empty.
He then ran all 250 entries in that selector. Seventy-three accept no ID class. The pattern he found is a US export-control table that dates to 1996. All 35 destinations in its Country Groups D:1 and D:5 fail, and no supported entry belongs to either group.
Approved, then locked out
Lubaretsi had already passed verification once. Persona checked his Georgian ID, OpenAI granted him trusted access, and he used it for defensive work. He builds open-source tooling for civil society groups, independent media and activists who face commercial and state-controlled spyware. On August 19, OpenAI Developers said a “technical issue” had deactivated Daybreak Blue, the tier that opens frontier models to verified defenders, for a limited set of users, and that they would need to verify again.
He resubmitted his ID and face scans eight times, using three different browser and device setups. At one point the flow stopped opening for about 24 hours. The result, with identifiers removed, shows "verification_status": "failed" in the same response as "individual_eligibility_status": "eligible" and "plan_eligible": true on a Pro plan. As of August 27, a thread about disappearing access on the OpenAI Community forum had 60 posts, and other previously approved users there reported the same mismatch between failed verification and confirmed eligibility, and the same lockout.
Roughly a week after the outage, the block moved earlier in the flow. A fresh account now fails as soon as Georgia is selected, with this screen:
Unable to verify. We are unable to verify identities in this country. Please select another country.
The access he lost was a modest upgrade to begin with. TAC does not remove guardrails, he writes, and verified users still trip aggressive classifiers. He describes the benefit as “slightly fewer classifier false positives.” OpenAI’s own figures, as Tech Journal reported them, point the same way. On its internal benchmark of exploit-chain, authentication-bypass and privilege-escalation requests, standard GPT-5.6 Sol completes 1.5%, Daybreak Blue access 2%, and the separately gated GPT-5.6-Cyber 95%. These are self-reported numbers and measure refusals only.
250 entries, one click each
For every entry, his recorder selected the country, pressed Select, captured Persona’s request and response, and saved screenshots before and after. The repository includes all 1,250 screenshots. In total, 177 entries support cyber verification and 73 do not. Of the 208 entries on OpenAI’s public ChatGPT availability list, 44 are blocked. Together those 44 markets have about 650 million people. Counting all 73 refused entries, the total is about 2.42 billion, with population data available for 61 of them.
Sanctions do not explain the list. Lubaretsi checked every entry against OpenSanctions. Only two of the 44 blocked ChatGPT markets matched an active US sanctions program under the audit’s filter. Georgia and Russia get the same answer from the selector, even though Georgia is on OpenAI’s ChatGPT list with no matched program, while Russia is absent from the list and matched three.
Thirty-five rows from 1996
The Commerce Department’s Bureau of Industry and Security puts 24 destinations in D:1, its national-security group, and 20 in D:5, which lists countries under US arms embargoes. Together that is 35 distinct entries. All 35 appear in OpenAI’s selector, and all 35 fail. D:1 accounts for the results that looked oddest in the audit: Georgia, Armenia, Azerbaijan, Moldova, Mongolia, all five Central Asian states, Vietnam, China and Russia.
The export table covers 35 of the 73 refusals. The other 38 include stranger cases: eleven French-associated entries fail while France passes, and Aruba fails while Curaçao passes. Lubaretsi leaves those unexplained. He argues that Georgia’s D:1 row “was never really written for independent Georgia” and was inherited from the Soviet Union.
The neighbouring groups do not trigger the gate. D:2 covers nuclear concerns, D:3 chemical and biological, and D:4 missile technology. Israel and Pakistan appear in all three, and both pass.
Who gets through?
Lubaretsi compared the gate with the 2024 World Cybercrime Index, which ranks countries by estimated concentration of cyber offenders. OpenAI accepts IDs from nine of the index’s top 15: Ukraine (2), the United States (4), Nigeria (5), Romania (6), the United Kingdom (8), Brazil (9), India (10), Ghana (13) and South Africa (14). It refuses Armenia (52), Georgia (66) and Azerbaijan (87), all three of which are ChatGPT markets.
Armenia is the sharpest case. On August 8, NVIDIA announced that the country had opened the CIS region’s largest AI factory. Firebird, the operator, plans more than 70,000 NVIDIA Rubin and Blackwell GPUs and 300 megawatts of AI capacity there by the end of 2027, and NVIDIA says it intends to invest in the company. Armenians cannot get verified for defensive use of the models that capacity will run.
What the rules ask for
Lubaretsi argues that the export rules OpenAI appears to be following do not require a blanket ban. Under 15 CFR § 742.4, the general policy for D:1 is to approve applications case by case when the items are for civilian use or would not significantly add to the destination’s military potential. The same section gives Kazakhstan and Mongolia “enhanced favorable consideration licensing treatment,” and both are blocked at the selector. The cyber-specific rule, 15 CFR § 740.22, restricts non-government users in D:1 and D:5 but makes exceptions for vulnerability disclosure, cyber incident response, deemed exports and specified favorable-treatment users. Georgia was already in D:1 in 2017, when the State Department approved a possible Javelin missile sale worth up to $75 million.
OpenAI’s own descriptions of the program are broader than the gate. Its published policy calls for clear, objective criteria such as KYC and identity verification, weighted by the user, trust signals, intended use and access level. The Daybreak overview adds risk considerations and the applicant’s ability to strengthen the cybersecurity ecosystem. OpenAI has also said verification includes trust signals such as device health, Biometric Update reported, and from September 1 individual members must authenticate with a hardware-backed passkey to keep frontier access. The country selector runs before any of these checks and reads only one thing: the country that issued the ID. Fouad Matin of OpenAI’s cyber preparedness team had put it differently in a post on X:
Trusted access isn’t limited to US/EU. We believe cyber capabilities should be broadly available with appropriate safeguards.
No retries, no appeals
At first, support treated Lubaretsi’s case as a technical problem. On August 25 it sent its final answer: “completing the flow does not mean that verification was approved,” and Trusted Access for Cyber verification “does not support retries or appeals.” Lubaretsi is asking OpenAI why membership in D:1 or D:5 became a categorical rule for individuals, why the supported-country list is only visible inside a biometric verification flow, and why a government ID is the only input the gate considers. The same week, on August 27, OpenAI joined more than 100 companies in an open letter asking frontier AI firms to give defenders access to their most capable models during major cyber incidents.
On August 23 he had emailed support his credentials, including FOI Security’s Democracy Tech Entrepreneur Award from the Copenhagen Democracy Summit 2026 and his earlier work at Transparency International Georgia. The reply came the next day: “Support cannot provide additional details about an individual verification decision or manually override its outcome. No further verification attempts or professional credentials are needed from you at this stage.”
Up next
Template sharing goes live in Grok Bot, Video
“Grok Bot now allows users to share bots as templates,” the SPACEXAI account posted on X. TestingCatalog had flagged that Templates tab in Settings weeks earlier as unreleased, with no purpose stated for the shared templates.
Related stories
- OpenAI opens Daybreak to Ukraine’s civilian cyber defense
- OpenAI halts some internal work with its unreleased Astra model
- OpenAI puts GPT-5.6-Cyber behind a new Daybreak Red tier
- Australia flags criminal penalties for harmful AI models
- Australia plans mandatory AI incident reporting
- OpenAI’s ad pixel fires before consent on 224 of 341 shops
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

