OpenAI Revises Security Rules: Possible Changes in Response to Competition | OpenAI API | OpenAI italiano | OpenAI stock | Turtles AI

OpenAI has updated its Preparedness Framework, introducing the ability to adapt security measures in response to high-risk AI releases from competitors. The changes include increased use of automated assessments and a new risk classification. However, former employees and critics are raising concerns about the company prioritizing safety over competitiveness.

Key Points:

• OpenAI may adjust security measures if competitors release high-risk AI without equivalent protections.
• The Preparedness Framework now uses more automated assessments to speed development.
• New risk categories include the ability of AI to self-replicate or evade controls.
• Former employees are rallying behind Elon Musk, criticizing the company’s prioritization of profits over safety.

OpenAI recently overhauled its Preparedness Framework, the internal system it uses to assess and mitigate risks associated with developing advanced AI models. Among the most notable changes, the company said it may adjust its safeguards if a competitor releases a high-risk AI system without equivalent protections. However, OpenAI stressed that such changes would only be made after a rigorous assessment of the risk landscape, public disclosure of the decision, and assurances that the overall risk of serious harm does not significantly increase.

The framework update also introduces a greater emphasis on automated assessments, which can help accelerate product development. While human testing has not been completely abandoned, OpenAI has developed a growing suite of automated assessments to keep up with a faster release cadence.

Additionally, the updated framework now classifies models based on new risk categories, including the ability to self-replicate, evade safeguards, or prevent shutdown. Models are assessed against two thresholds of capability: “high,” for those that could amplify existing paths to serious harm, and “critical,” for those that introduce unprecedented new paths to such harm. Systems that reach high capacity must have safeguards in place that sufficiently minimize the associated risk before they are deployed, while those with critical capacity require such safeguards during development.

However, these changes have drawn criticism. According to the Financial Times, OpenAI gave testers less than a week to perform safety checks on a major upcoming model, a shorter timeframe than previous releases. Additionally, many of the safety tests are now being conducted on older versions of models rather than those released to the public.

Concerns have been further fueled by a brief filed by 12 former OpenAI employees in Elon Musk’s lawsuit against the company. They argue that OpenAI would have been encouraged to further cut security costs if it completed its planned corporate restructuring, which aims to transform the organization into a more profit-oriented entity.

In response to the criticism, OpenAI says it is keeping its safety measures at a more protective level and does not make policy changes lightly. The company emphasizes the importance of a scientific approach to risk management and is committed to engaging independent third parties to verify the technology and provide feedback, as well as working with external parties and internal teams to monitor misuse and misalignment risks of its models in the real world.

The debate over OpenAI’s priorities between safety and competitiveness continues, with observers and former employees raising questions about the company’s direction in developing advanced AI.