The most recent governance frameworks from OpenAI provide enterprise leaders with a well-defined template for scaling their safe and compliant use of AI around the world.
There have been developments on large language models which require a commercially robust architecture approach. OpenAI recently launched its Frontier Governance Framework (FGF), which outlines its risk assessment and mitigation process within the organization.
The Frontier Governance Framework is designed based on the guidelines set by both the General-Purpose AI Code of Practice in the European Union and California’s Transparency in Frontier AI Act. The publication acts as an extremely practical guide showing how companies’ systems and processes should be designed to safely utilize machine learning capabilities.
However, the first step in developing this into a business strategy is to understand the definition of threat types. Systemic risk refers to foreseeable material risks of harm. In concrete terms, it refers to situations whereby a model is associated with more than 50 deaths or $1 billion worth of damages.
Although these examples might lie on the fringes of possibility, the creation of them means that deployment teams can take measures that safeguard against such a possibility. Once the parameters have been set out, companies can allocate their compute hours and engineering effort towards maintaining monitoring and auditing for compliance.
Also Read : Display Ads Are Moving into Google’s AI-Powered Demand Gen
Using risk assessments in tiers for internal systems
Threats in OpenAI can be classified in different categories including cyber offense, chemical, biological, radiological, and nuclear (CBRN), harmful manipulation, and loss of control.
The threat classification scheme is based on the use of different risk levels depending on the model capabilities. For instance, Tier 3 level of cyber offense risk indicates that the model enhanced by tooling can recognize, identify, and create effective zero-day exploits of all levels of complexity and in numerous well-protected environments without any human involvement.
In CBRN threat classification, a model in Tier 3 allows creating an effective new threat vector that can be compared to a CDC Category A biological agent or autonomously completing the cycle of creating a regulated biological threat. The problem is not only in the dangerous consequences of these actions but also in the possibility to use risk levels to define specific boundaries of activities in which a particular model may be involved. It means that the security team can understand when some coding assistance should be subject to more thorough monitoring.
The other type of threats related to harmful manipulation refers to the intentional distortion of human behavior and actions.
According to OpenAI, this domain is still experimental and should be handled via system-level mitigation, such as post-deployment monitoring, and not pre-deployment assessments. In consumer-facing industries, this means that marketing automation systems leveraging language models only need content classifiers to produce objective communication in real-time.
Dealing with the threat posed by the lack of human ability to control or terminate the operation of the system, this risk vector is classified as a loss of control. The Tier 2 model in this risk vector would exhibit the capacity to consistently avoid detection in multiple assessment methods, such as the avoidance of chain of thought monitoring.
The Tier 3 model represents the ability of the system to be superior to the most experienced humans in completing the majority of complex tasks and be able to function independently for a significant amount of time. Moreover, the Tier 3 system would exhibit advanced situational awareness and stealth where even monitoring the model and its chain of thought is unable to reliably detect evasion of control.
Also Read : OpenAI Opens Singapore AI Lab as Infocomm Media Development Authority Updates AI Framework
Integration issues and information security issues
OpenAI ensures that their internal security adheres to ISO standards such as ISO 27001, 27017, 27018, and 27701 in addition to SOC 2 Type II. To safeguard unpublished model weights, OpenAI makes use of encryption for stored and transferred data, multi-factor authentication, and strict multi-party approval protocols. Employees receive consistent training sessions and sandboxing with restricted egress is used for model execution.
As companies replicate this model, they develop a security framework internally.
For integration of models with private corporate data sets, engineers find themselves having to use models like Retrieval-Augmented Generation and dense vector databases. In order to secure these databases against attempts at adversarial prompting and data extractions, specific computation costs must be incurred.
Before any request goes to the vector database and retrieved context is returned, every API call will have to pass through the security classifier to ensure that the response generated from the query is accurate. In order to bridge between modern cloud-based governance structure of artificial intelligence and the older mainframe data silos, bespoke heavy encryption middleware will need to be constructed by engineers.
Compliance and Incident Response of Ecosystems
For the accuracy of the risk baseline, OpenAI seeks feedback from outside domain experts as well as independent third parties who assess the safeguards of models reaching new risk tiers.
Similarly, CDOs within an organization can engage external audit retainers to ensure that the localized deployment of models stays within acceptable risk thresholds.
From the point of view of the regulatory framework, the external reporting determines the operational rhythm. For instance, OpenAI publishes findings of the mitigation process into a Safety and Security Model Report. The report will be updated at least every six months for the most capable models, as per the provisions of the EU AI Act.
Changes that affect the capabilities of a model post-training or increase risks associated with integration into internal systems are regarded as required updates to the reports. It is OpenAI Ireland Limited that ensures compliance to EU regulations. Meanwhile, OpenAI OpCo LLC takes care of TFAIA compliance in the USA.
OpenAI applies an AI Safety Incident Response Plan called AIRP to react swiftly to software abnormalities. The plan includes procedures to be taken for triaging, investigating, and external notification regarding major safety incidents.
Abnormal situations are detected through automated detection, employee reporting, and user notifications. Following that, a response team investigates the root causes and impact of the incident and takes measures to mitigate and contain the incident. Such procedures are easy to emulate for enterprise leaders. One should create their own response unit able to correct the behavior of the anomalous API.
OpenAI has different people in leadership positions that can initiate updates to the framework. The Head of Safety Systems, CISO, and General Counsel could all initiate changes to the safety management system. An annual Framework Assessment is carried out at least once per year.
