AI models going rogue and hacking the internet? That’s “the stuff of science fiction”, said The New York Times’ Katrin Bennhold. “Or at least it was until recently.”
The UK’s AI Security Institute announced on Tuesday that two AI agents it was evaluating – one powered by Anthropic’s Mythos 5 and the other by OpenAI’s GPT 5.6 Sol – had taken “sustained, unsanctioned action directed at real people and organisations”. In the most serious incident, the Mythos 5 agent created “fake online identities”, tried to deceive a software engineer, and attempted “to insert malicious code into a open-source project”.
Only a few days ago, Open AI and Anthropic revealed similar testing mishaps, in what was a “big moment” for the world, Nate Soares, of the non-profit Machine Intelligence Research Institute, told The New York Times. “These AIs were committing cybercrimes a human would be strongly punished for”. It is, “in a sense”, artificial intelligence’s “first felony”.
The Week
Escape your echo chamber. Get the facts behind the news, plus analysis from multiple perspectives.
SUBSCRIBE & SAVE
Sign up for The Week's Free Newsletters
From our morning news briefing to a weekly Good News Newsletter, get the best of The Week delivered directly to your inbox.
From our morning news briefing to a weekly Good News Newsletter, get the best of The Week delivered directly to your inbox.
Latest Videos FromThe Week
“The ‘robots could kill us all’ argument” has long been dismissed “as hysteria or even calculated hype”, said the NYT’s Bennhold. But fears that frontier AI models were “getting too good too effective at exploiting software vulnerabilities” have “now become reality”. There needs to be “intense debate about the dangers of AI spiralling out of human control and posing a threat to humanity”.
AI engineers “and even their vainglorious bosses, are getting freaked out by the capabilities they are handling and worry about unchecked proliferation”, said Rafael Behr in The Guardian. The obvious historical analogy is nuclear fission, which could have been “harnessed benignly” or “deployed aggressively in warheads”. But now “there are multiple Manhattan Projects all frantically competing for market share, fuelled by trillions of dollars of debt” and “with the rest of the US economy as collateral”.
Negotiating international agreement on a safe way forward for frontier AI is going to be tricky, said The Economist, when it is “increasingly clear” that the US and China will “restrict access to their models for economic and strategic reasons”.
Chinese officials are already worried “about the cyber-capabilities of Mythos” and “its potential to be used as an offensive weapon, said Bloomberg. And, for the US and its allies, there is deep concern that their geopolitical adversaries “and the criminals in their orbit, will soon wield the same offensive tools”, said Michael Sulmeyer, former US assistant secretary of defence for cyberpolicy, on Foreign Affairs. With China, in particular, “already racing to build and acquire these AI capabilities”, the US may only have “nine to 12 months to protect its critical infrastructure from AI-powered attacks”.
What next?
Any kind of “consistent approach” to global AI regulation has “failed to materialise”, said Sky News. The Trump administration is finalising a voluntary framework under which US AI labs would submit models for federal safety testing before public release. But this latest “alarming behaviour” is “likely to ignite fresh calls” from lawmakers and Silicon Valley for “more rigorous regulation”, said Politico’s John Sakellariadis.
The recent incidents also raise the question of who bears responsibility when AI systems break the law. If those hacking attempts had been “human-originated, they would lead to clear and vigorous prosecution”, ethical hacker and cybersecurity expert Marc Rogers told Politico. “It’s time for a serious discussion about updates to existing computer security law.” Until that happens, said Andrew McDonald on Politico London Playbook, “this kind of thing is going to become more common, and quickly”.
