Two rival labs, almost signing an agreement to undermine each other.
According to media reports, earlier this year OpenAI and Anthropic came close to reaching a binding agreement that would allow both sides to test each other’s models, looking for vulnerabilities and hidden safety risks. Under the disclosed terms, each side would gain programming access to the other’s commercial models to conduct vulnerability testing, and would be prohibited from retaining any of the other side’s data.
The scope is drawn very clearly: unpublished systems are not included. In other words, what can be checked is what has already been sold, while the parts that haven’t been released yet stay behind closed doors. This arrangement hands a portion of the verification power to peers, and turns peers’ judgments into a form of external oversight.
It’s not clear whether the agreement was ultimately signed. Handing the most sensitive checks to a competitor doesn’t require goodwill so much as both sides believing the other will follow the rules. For two companies that are mutual rivals, the scarcest resource has always been trust: both are picking apart the other while simultaneously refusing to let the other see its bottom line.
In the same week, executives at one company publicly questioned whether one of them could still make it to an IPO, while outright excluding the other. Industry trust, it seems, is only brought out when there’s something to prove.
The ceiling for security cooperation is usually written in the few lines of text within the disclosure scope.
#人工智能 #安全
According to media reports, earlier this year OpenAI and Anthropic came close to reaching a binding agreement that would allow both sides to test each other’s models, looking for vulnerabilities and hidden safety risks. Under the disclosed terms, each side would gain programming access to the other’s commercial models to conduct vulnerability testing, and would be prohibited from retaining any of the other side’s data.
The scope is drawn very clearly: unpublished systems are not included. In other words, what can be checked is what has already been sold, while the parts that haven’t been released yet stay behind closed doors. This arrangement hands a portion of the verification power to peers, and turns peers’ judgments into a form of external oversight.
It’s not clear whether the agreement was ultimately signed. Handing the most sensitive checks to a competitor doesn’t require goodwill so much as both sides believing the other will follow the rules. For two companies that are mutual rivals, the scarcest resource has always been trust: both are picking apart the other while simultaneously refusing to let the other see its bottom line.
In the same week, executives at one company publicly questioned whether one of them could still make it to an IPO, while outright excluding the other. Industry trust, it seems, is only brought out when there’s something to prove.
The ceiling for security cooperation is usually written in the few lines of text within the disclosure scope.
#人工智能 #安全
