OpenAI’s Third-Party AI Safety Rules Lacks Enforcement, Critics Say

OpenAI published a detailed set of priorities and principles on Tuesday designed to govern its third-party model assessors. While industry observers agree the framework establishes useful benchmarks for independent evaluation, critics emphasize that the document lacks enforceable controls to guarantee true safety compliance or accountability.

The Illusion of Independent Scrutiny Under Corporate Terms

The core tension of OpenAI’s new governance framework lies in its self-imposed boundaries. According to the published document, third-party access is explicitly granted within the bounds of legal, security, and intellectual property constraints. This means the lab retains ultimate authority to decide the scope of its own scrutiny.

Frank Dickson, principal analyst at Dickson Research, pointed out that the document functions primarily as a code of conduct for the assessors rather than a binding constraint on OpenAI itself. The text specifies that publication of findings happens only after labs receive a reasonable period to remediate issues. That mirrors the grace-period logic that previously drew criticism elsewhere in the tech sector, such as Google delaying Gemini hacking disclosures, now formalized as a governing principle.

Pieter Arntz, a malware intelligence researcher at Malwarebytes, noted that while setting rules of engagement is a positive step, the document says nothing about an actual enforcement process. It does not compel OpenAI to submit to a specific scope, publish adverse findings, or alter deployment decisions. Its value entirely hinges on whether the lab accepts genuinely inconvenient oversight.

Commercial Pressures Versus Original Safety Mandates

For enterprise chief information officers weighing whether to trust proprietary models, this half-measure does little to build confidence. Jason Andersen, principal analyst at Moor Insights & Strategy, described OpenAI as a split-brain company caught between its original non-profit research mission and aggressive commercial imperatives. Over the past two years, those two internal cultures have struggled to align, resulting in compliance frameworks that read more like corporate legalese than ironclad safety guarantees.

Valence Howden, an advisory fellow at Info-Tech Research Group, added that OpenAI retains control over assessor access, enabling the lab to filter what is provided and impede the gathering of direct evidence whenever material or classified data is involved. Because the company controls the reporting and retains redaction rights, true transparency remains heavily mediated.

Analyst / Source	Organization	Core Critique of the Assessment Framework
Pieter Arntz	Malwarebytes	Lacks mechanisms to compel scope compliance, adverse finding publication, or deployment changes.
Valence Howden	Info-Tech Research Group	OpenAI retains control over access and redactions, limiting direct evidence gathering.
Frank Dickson	Dickson Research	Functions as a code of conduct for assessors, embedding remediation grace periods rather than accountability.

Moving From Risk Assessment to Enforceable Accountability

Despite the widespread skepticism over current enforcement teeth, some risk management professionals view the release as an initial baseline from which a more robust governance model can evolve.

Samantha Gloede, global head of risk services at KPMG, argued that the conversation must inevitably shift from mere assessment to structural accountability. Independent testing matters, but long-term enterprise trust depends on how organizations respond when material risks are identified, independently validated, and integrated into deployment decisions.

Edna Conway, an executive advisor at Acceligence, emphasized that the true substance of this effort will depend on subsequent phases. Making labs accountable to the assessment process requires answering fundamental questions about who determines when an evaluation is required, what constitutes sufficient access, and how critical findings are reported when the lab and the assessor disagree.

Until those operational parameters are codified into binding, external oversight, OpenAI’s new priorities remain an initial sketch rather than a finished blueprint for safety.