NHS England is pulling its open-source software from the internet because of fears around computer-hacking AI models like Mythos. A decision by NHS England to withdraw open-source code created with UK taxpayer funds because of the risk posed by computer-hacking AI models is attracting growing backlash. Last month, Mythos, an AI created by technology firm Anthropic, was widely reported to be capable of discovering flaws in virtually any software, potentially allowing hackers to break into systems running it. NHS England has now told staff that existing and future software must be pulled from public view and kept behind closed doors by 11 May because of this risk. The decision goes against the NHS service standard, which requires that staff make any software they produce open-source so that tools can be built upon, improved and used without the need for duplicated effort.

NHS England is hurriedly withdrawing all the software it has written from public view because of the perceived risk of hacking from cutting-edge artificial intelligence. Security experts say the move is unnecessary and counterproductive. Software produced by the National Health Service has previously been made open-source and listed on GitHub because it is created with public money. This allows other organisations to build upon it and make better services more cheaply without duplicating effort. But NHS England has issued new guidance to staff, which has been shared with, that demands existing and future software be pulled from public view and kept behind closed doors.

Do you need to worry about Mythos, Anthropic's computer-hacking AI? A powerful AI kept from public access because of its ability to hack computers with impunity is making headlines around the world. But what is Mythos, does it really represent a risk and might it even be used to improve cybersecurity? Anthropic's Project Glasswing aims to improve online security The past few weeks have brought apparently alarming news of Mythos, an AI that can identify cybersecurity flaws in a matter of moments, leaving operating systems and software vulnerable to hackers. The cybersecurity community is now beginning to get a better sense of how Mythos may change the face of cybersecurity - and not necessarily for the worse.

'The US government's embrace of Anthropic marks a shift.' 'The US government's embrace of Anthropic marks a shift.' The Guardian view on Anthropic's Claude Mythos: when AI finds every flaw, who controls the internet? A nthropic announced its latest AI model, Claude Mythos, this month but said it would not be released publicly, because it turns computers into crime scenes. The company claimed that it could find previously unknown "zero-day" flaws, exploit them and, in principle, link these weaknesses in order to take over major operating systems and web browsers . Mythos did so autonomously, writing code and obtaining privileges.

'Releasing a marketing post with purposely vague language that obscures evidence brings into question if they are trying to garner further investment without scrutiny,' one scientist said. 'Releasing a marketing post with purposely vague language that obscures evidence brings into question if they are trying to garner further investment without scrutiny,' one scientist said. 'Too powerful for the public': inside Anthropic's bid to win the AI publicity war This week, the AI company Anthropic said it had created an AI model so powerful that, out of a sense of overwhelming responsibility, it was not going to release it to the public. The US treasury secretary, Scott Bessent, summoned the heads of major banks for a chat about the model, Mythos. The Reform UK MP Danny Kruger wrote a letter to the government urging it to " engage with AI firm Anthropic whose new frontier model Claude Mythos could present catastrophic cybersecurity risks to the UK".

'Lethal cyber-attacks are thankfully rare. But a new AI release could change that.' 'Lethal cyber-attacks are thankfully rare. But a new AI release could change that.' Anthropic's new AI tool has implications for us all - whether we can use it or not Claude Mythos's apparent superhuman hacking abilities are alarming experts as the Trump administration remains blinded by hostility I n June 2024, a cyber-attack on a pathology services company caused chaos across London's hospitals. More than 10,000 appointments were cancelled. Blood shortages followed and delays to blood tests led to a patient's death . Lethal cyber-attacks like this are thankfully rare.

New Jersey man's chilling'cancer map' fuels fears of poisoned neighborhood with 41 cases and counting Three stocks are high as a kite after Trump's wild executive order as investors rush to cash in New'Hollywood dose' pill: A-listers hooked on'youth elixir' that dermatologists say is anti-ageing, shrinks pores, smooths wrinkles... and even banishes rosacea Days after we got engaged, the love of my life told me he'd killed a man and buried him in a bog. I reported him to police... but then I made this irreversible mistake Papa John's under fire for an outrageous message now printed on all pizza boxes Iran vows to put'new cards on the battlefield' after Trump breaks ceasefire as Vance travels to Pakistan for peace talks before deadline ends TODAY NASA's return of humans to the Moon in 2028 faces alarming setback California coffee farmers nearly escaped death before'tragic accident' as autopsy reveals disturbing new details How to lose weight when perimenopause sabotages your metabolism: I'm a PT but when I hit 46, I piled on the pounds overnight. Australia has spoken: Report reveals what everyone is thinking about Prince Harry and Meghan Markle's Australia tour Humiliating moment runner celebrates winning marathon... only to be pipped at the line by rival in brutal finish How prophet of extreme Mormon cult who had 20 wives - some aged just 10 - is now spreading evil from prison, as woman who bravely exposed him reveals new threat Netflix doc missed and'sister brides' still under his thrall Even Cameron Diaz admits she's a dirty mess. I'll get hate for saying it, but we're all thinking the same thing about THAT wrinkled forehead: CAROLINE BULLOCK Two high school sweethearts survived the Columbine High School massacre. Months later, they were gunned down in a Subway on Valentine's Day in a crime that remains unsolved AI-pocalypse: Anthropic sparks fears after developing a bot that's'too dangerous to release to the public' Anthropic has sparked fears after revealing that it has developed an AI bot deemed too dangerous to release to the public.

Language model alignment methods, such as reinforcement learning from human feedback (RLHF), have led to impressive advances in language model capabilities, but existing techniques are limited by a widely observed phenomenon known as overoptimization, where the quality of the language model plateaus or degrades over the course of the alignment process. Overoptimization is often attributed to overfitting to an inaccurate reward model, and while it can be mitigated through online data collection, this is infeasible in many settings. This raises a fundamental question: Do existing offline alignment algorithms make the most of the data they have, or can their sample-efficiency be improved further? We address this question with a new algorithm for offline alignment, $\chi^2$-Preference Optimization ($\chi$PO). $\chi$PO is a one-line change to Direct Preference Optimization (DPO; Rafailov et al., 2023), which only involves modifying the logarithmic link function in the DPO objective. Despite this minimal change, $\chi$PO implicitly implements the principle of pessimism in the face of uncertainty via regularization with the $\chi^2$-divergence -- which quantifies uncertainty more effectively than KL-regularization -- and provably alleviates overoptimization, achieving sample-complexity guarantees based on single-policy concentrability -- the gold standard in offline reinforcement learning. $\chi$PO's simplicity and strong guarantees make it the first practical and general-purpose offline alignment algorithm that is provably robust to overoptimization.

Supervised machine-learning models boast remarkable predictive capabilities. But can you trust your model? Will it work in deployment? What else can it tell you about the world? Models should be not only good, but also interpretable, yet the task of interpretation appears underspecified.

Supervised machine learning models boast remarkable predictive capabilities. But can you trust your model? Will it work in deployment? What else can it tell you about the world? We want models to be not only good, but interpretable. And yet the task of interpretation appears underspecified. Papers provide diverse and sometimes non-overlapping motivations for interpretability, and offer myriad notions of what attributes render models interpretable. Despite this ambiguity, many papers proclaim interpretability axiomatically, absent further explanation. In this paper, we seek to refine the discourse on interpretability. First, we examine the motivations underlying interest in interpretability, finding them to be diverse and occasionally discordant. Then, we address model properties and techniques thought to confer interpretability, identifying transparency to humans and post-hoc explanations as competing notions. Throughout, we discuss the feasibility and desirability of different notions, and question the oft-made assertions that linear models are interpretable and that deep neural networks are not.