Inference Investigation (External Handoff)

Claim investigated: The specific capabilities OpenAI provides to defence and intelligence agencies are classified or protected as trade secrets. Entity: OpenAI Original confidence: inferential Result: STRENGTHENED → SECONDARY Source: External LLM (manual handoff)

Assessment

The claim is strongly supported by OpenAI’s confirmed defence contracts (CDAO awards for 'frontier AI projects') and the inherent sensitivity of such work. Defence and intelligence applications of frontier AI typically involve classified information, while OpenAI’s proprietary models are legally protected as trade secrets. The established documentation gaps for recent contracts and the use of OTA/subcontractor vehicles further reinforce that specific capabilities are withheld from public disclosure.

Reasoning: Primary facts confirm OpenAI holds defence contracts (CDAO awards), while secondary facts establish recent, incompletely documented procurement via OTA/subcontractor arrangements. Frontier AI for national security is routinely classified, and OpenAI’s models are explicitly protected as trade secrets under US law. The combination of confirmed defence engagement, typical classification of such applications, and trade secret protections for proprietary AI models creates a compelling evidentiary basis that specific capabilities are not publicly disclosed, elevating the inference to well-supported secondary confidence.

Underreported Angles

• The classification levels (Confidential, Secret, Top Secret, SAP) applied to OpenAI’s defence contracts, which determine disclosure restrictions
• Whether OpenAI has formal Special Security Agreements (SSAs) with DoD or intelligence agencies governing access to classified capabilities
• The specific model versions (e.g., custom fine-tunes of GPT-4) deployed for defence applications, which may differ from public releases
• Data access protocols allowing defence/intelligence agencies to input classified information into OpenAI models without compromising security
• Deployment architectures (Azure Government, classified clouds, air-gapped systems) that enable secure use of OpenAI capabilities in national security contexts
• The specific use cases (surveillance, targeting, cyber operations, intelligence analysis) for which OpenAI’s models are being employed
• Any export-controlled or ITAR-regulated modifications to OpenAI’s models for defence applications
• Trade secret assertions OpenAI has made in court filings or contracts regarding its defence-related AI capabilities
• The security clearance status of OpenAI employees working on defence contracts and how this affects capability disclosure
• Whether OpenAI’s defence work involves custom-trained models or standard models with restricted access and usage parameters

Public Records to Check

• USASpending: "OpenAI" AND (CDAO OR "Chief Digital and Artificial Intelligence Office" OR DoD OR "Department of Defense") Contract descriptions may reveal redacted or generic capability references, confirming classification or trade secret withholding

• DoD contract announcements: "OpenAI" AND (frontier OR AI OR "artificial intelligence" OR model OR capability) Official announcements might provide unclassified summaries of capabilities, or their absence would support non-disclosure

• FOIA: "OpenAI" AND (CDAO OR DoD OR "defense" OR "intelligence" OR contract OR capability) FOIA releases could reveal contract details or redactions that confirm classification or trade secret protections

• Congressional hearings: "OpenAI" AND (DoD OR defense OR intelligence OR CDAO OR capability) Testimony or Q&A might reference capabilities at a level of detail not available in contract databases

• Court records: "OpenAI" AND (trade secret OR "proprietary" OR model OR capability OR defense) Litigation (e.g., against former employees) may reveal how OpenAI designates its defence-related AI as trade secrets

• Patent databases: "OpenAI" AND (defense OR intelligence OR government OR military OR "national security") Patent filings might reveal defence-related AI capabilities, or their absence would support trade secret protection

• SEC EDGAR: Microsoft AND (OpenAI OR Azure Government OR defense OR intelligence OR AI) Microsoft’s filings may disclose material details about OpenAI’s defence work through Azure Government contracts

• SBIR/STTR databases: "OpenAI" AND (DoD OR "Department of Defense" OR intelligence OR research) Would capture any defence research contracts that might describe capabilities in unclassified terms

• National Laboratory contracts: "OpenAI" AND (Los Alamos OR Sandia OR Livermore OR "national laboratory") National labs often work on classified AI projects and their contracts may reference OpenAI capabilities

• LDA: "OpenAI" AND (DoD OR defense OR intelligence OR CDAO OR capability) Lobbying disclosures might reference defence capabilities in discussions with Congress or agencies

Significance

CRITICAL — The non-disclosure of specific frontier AI capabilities deployed in national security creates a critical accountability gap regarding oversight of potentially world-altering technology, its applications in defence and intelligence, and the alignment between corporate trade secret interests and public safety concerns