If Anthropic is kind of like a kid at a Halloween party with a bowl full of spaghetti, telling its blindfolded friends the spaghetti is brains, European cyber agency leaders are like kids who are mad that they still haven’t gotten to stick their fingers in the brain bowl, according to a report by Politico.
That publication reached out to eight agencies in continental Europe, and they sound a bit jealous of U.S. tech companies, along with the U.S. and U.K. governments, which have all gotten a nice taste of the apparent super-hacking capabilities of Anthropic’s unreleased model, Claude Mythos Preview.
Through a spokesperson named Job Holzhauer, the Dutch cybersecurity agency told Politico, “the actual impact of the vulnerabilities found is difficult to verify without technical details.”
British authorities, meanwhile, have touched the spaghetti. I wrote yesterday that Claude Mythos Preview, has scared officials in the U.K. lavishly, and they’ve gotten all the gory details. Kanishka Narayan, the U.K.’s AI minister says the UK-based AI Security Institute has had sufficient access to test the model in some capacity, and it seems the test unearthed something significant. “We’ve taken action based on our findings,” Narayan wrote on X.
It’s worth noting that Politico’s report gives the impression that the eight European agencies it spoke to were disappointed with access so far, but some had apparently been allowed some unknown type of limited preview. Surprisingly, Germany—home of the largest economy in Europe—had “entered into conversations with Anthropic about Mythos,” in Politico’s phrasing, but it “had not yet been able to test the model.”
And intriguingly, an AI researcher named Laura Caroli opined to Politico that it had been possible for the EU to be “sidelined,” because the model is still unreleased. Apparently if it were on the open market, it would be bound by EU laws, and certain legal obligations would kick in.
