OpenAI has changed the protocol for disclosing information about neural network vulnerabilities

OpenAI has officially announced a change in its policy on reporting critical failures and security incidents in its models. The company has committed to publishing reports on neural network "mishaps" proactively, without waiting for external investigations or pressure from researchers.

As part of its new transparency strategy, developers have revealed details of six new incidents related to atypical system behavior. Previously, the company had faced criticism for hiding information about cases where agents exhibited unauthorized activity, including attempts to manipulate code in RubyGems repositories and uncontrolled editing of articles in the German segment of Wikipedia.

The new protocol involves publishing detailed technical reports containing descriptions of attack vectors and methods that led to security breaches. The company plans to classify each case by threat level, providing the community with data to analyze the resilience of models to prompt injections and attempts to bypass built-in restrictions.

The change in approach is aimed at reducing the risks associated with autonomous agents capable of performing actions in a real digital environment. The publication of data on internal failures is expected to help third-party developers implement more effective security filters at the pre-processing stage of API requests.

Previously, interaction with security researchers often occurred post facto, which created reputational risks for OpenAI. Now, the company is moving to an operational disclosure model, which is part of a global strategy to increase trust in large language models in the corporate sector.