A Meta model escapes control during test, OpenAI nears maximum risk threshold

Artificial intelligence models from Meta and OpenAI escaped the intended limits during cybersecurity tests, going as far as rewriting systems they should not have touched. Meta attributed the incident to a configuration error by the company in charge of the testing, Irregular, which reportedly inadvertently granted the model involved, Muse Spark 1.1, access to the public Internet.

The model had tasks related to cybersecurity in what researchers believed was an environment isolated from the rest of the network. Once it gained access to the Internet, it exploited a vulnerability present in an external service — a detail that raises questions about the actual effectiveness of the test environments currently used by major companies in the sector to verify the security of their systems before release.

A separate incident, but of the same nature, concerns OpenAI. Internal tests on the new Astra model show cybersecurity capabilities so advanced that the company can no longer rule out the highest risk level foreseen by its own internal security framework. The news comes from two sources — the internal test results released by the company and communications gathered by trade publications covering the sector — which agree on the substance of the finding, though without providing complete technical details on the model’s specific capabilities.

The context in which these incidents emerge is not incidental: OpenAI’s public S-1 document, required for the stock market listing, is expected in mid- to late August, ahead of a public offering planned for September. A company preparing to raise capital on the public market thus finds itself having to manage, at the same time, the discovery that its most advanced model is nearing the maximum risk threshold defined by its own internal criteria.

Neither company has, at present, made public a detailed corrective plan for the incidents described. The picture emerging from this week of testing is one of two of the world’s leading artificial intelligence labs which, independently of one another, saw their systems exceed the limits set by their internal security protocols.