Scott Aaronson reported that some AI companies have started to discreetly check whether your latest internal models can break important cryptographic protocols and primitives. The theoretical computer scientist attributed the information to his own sources, without announcing a successful breach.
The report appears in an update to his speech from October 7. Aaronson does not identify the companies involved, the algorithms tested, or the results obtained. The point that can be ascertained through the text is therefore limited: the author reports the start of an investigation into the capabilities of the models.
For Aaronson, any success would make it timely anticipate its consequences before others arrive at the same discovery. The reasoning remains conditional: the post offers no elements to establish that that ability already exists.
Confidential experiments and public tests are two distinct steps
The distinction matters when reading the news. Trying to violate a protocol is not the same as having violated itand the confidentiality of an experiment does not demonstrate its success. Without verifiable results, attributing to these models a concrete ability to compromise encryption would mean going beyond what has been reported.
To evaluate a possible announcement they would need at least the precise objective, the conditions of the experiment and a verification of the result. They are the elements that would allow us to distinguish a relevant discovery from a demonstration with limits that are too stringent to support general statements on the safety of systems.
The question posed by the tests deserves attention, but the answer remains open. The available information does not allow us to indicate compromised protocolsmeasure the effectiveness of experiments or attribute responsibility to individual producers. Technical evidence will be needed: until then, the news concerns a research activity reported by Aaronson and its possible developments.
