AI company Anthropic withholding advanced model over security risks
Anthropic withholds advanced Claude Mythos model from public release, limiting access to ~40 partners via Project Glasswing to let defenders patch vulnerabilities before similar AI capabilities proliferate.
Objective Facts
Anthropic is rolling out a preview of its new Mythos model only to a handpicked group of tech and cybersecurity companies over concerns about its ability to find and exploit security flaws, with Anthropic so worried about the damage Mythos could cause that it's refusing to release it publicly until there are safeguards to control its most dangerous capabilities. Mythos Preview is "extremely autonomous" and has sophisticated reasoning capabilities that give it the skills of an advanced security researcher, and can find "tens of thousands of vulnerabilities" that even the most advanced bug hunter would struggle to find. The company said on Tuesday it was rolling out Mythos through an invitation-only initiative called Project Glasswing, restricted to defensive cybersecurity work and limited to around 40 organizations, aimed at giving cyber defenders a head start on securing some of the world's most critical software systems from the looming security risks posed by advanced AI. Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell convened an urgent meeting with bank CEOs this week to warn of cyber risks posed by Anthropic's latest AI model and make sure that banks are implementing appropriate safeguards to protect their systems.
Left-Leaning Perspective
Progressive outlets and AI governance advocates have emphasized concerns about private corporate control of powerful technology. Fortune reported that some security experts and software developers committed to open-source software argue the world would be safer if Mythos were released so that every defender, not just Anthropic's chosen partners, could use it, with Jonathan Iwry from the Wharton Accountable AI Lab stating "Whatever the right judgment call is, the most striking aspect of this situation is how reliant we are on the judgment of a handful of private actors who arent accountable to the public". AI researcher Gary Marcus acknowledged Anthropic's restraint in not releasing the model but warned that "some of their competitors (such as OpenAI and xAI) might well not" show such caution, and argued that "without any government oversight on what can be released, we are entirely at the mercy of individual CEOs, some of whom have decidedly not earned our trust". Left-leaning critics see this moment as revealing broader governance gaps—Anthropic gets credit for caution, but the fact that one company can unilaterally decide who accesses world-altering technology underscores the need for democratic oversight rather than relying on corporate virtue. Progressively-minded voices also note Anthropic's courageous stance against the Trump administration. USA Herald commentary noted that "Anthropic has taken a principled position in resisting what it perceives as potential misuse of its models, even when that stance places it at odds with powerful government interests," helping to define "the emerging doctrine of A.I. governance, where private developers may assert limits on how their technology is used, even by state actors". Anthropic's public refusal to allow Claude for autonomous weapons and mass surveillance, which triggered the Pentagon's "supply chain risk" designation, appeals to those concerned about militarization of AI. What left-leaning coverage generally omits or downplays is the possibility that Anthropic's restricted access strategy might genuinely be the best available option given the speed of AI capability diffusion, or that the company may face real technical constraints (computational resources, safety validation requirements) that limit its ability to deploy at scale immediately. Progressive coverage tends to frame access restriction primarily as a power/governance problem rather than engaging with the technical and market-timing dilemmas.
Right-Leaning Perspective
Conservative and market-oriented critics have questioned whether Anthropic's withholding is genuinely about safety or primarily about business strategy. TechCrunch's Tim Fernholz reported that limiting release to big organizations "creates a flywheel for big enterprise contracts, while making it harder for competitors to copy their models using distillation, a technique that leverages frontier models to train new LLMs on the cheap". Fernholz suggested that "Anthropic may have found a clever approach to protecting the internet — and its bottom line". Some analysts have dismissed these cautious, limited releases as more about marketing and creating hype around new models, rather than purely safety-driven decisions. Right-leaning and skeptical observers also question the technical claims themselves. Tom's Hardware noted that "Many of the 'thousands' of bugs and vulnerabilities it found are in older software, or are impossible to exploit," and that "Under the subheading, 'and several thousand more,' Anthropic also states that it can't actually confirm that all of the thousands of bugs Mythos claims to have found are actually critical security vulnerabilities". TechCrunch reported that "it's not clear that Mythos is actually the be-all and end-all of cybersecurity models," noting that "Aisle, an AI cybersecurity startup, said it was able to replicate much of what Anthropic says Mythos accomplished using smaller, open-weight models," with Aisle's team arguing "there is no single deep learning model for cybersecurity, but instead depends on the task at hand". Right-leaning coverage tends to downplay or omit the genuine technical arguments that larger, more capable models may pose qualitatively different risks than smaller open models—arguments made by security researchers who have tested both. It also frames Anthropic's Pentagon dispute primarily through a lens of corporate overreach rather than engaging with questions about whether companies should have input on military applications of their technology.
Deep Dive
Anthropic's decision to restrict Mythos access sits at the intersection of three major tensions in 2026 AI governance: the speed of capability diffusion, the role of private companies in making national security decisions, and the question of whether safety concerns are genuine or primarily commercial. Anthropic argues that models with advanced cybersecurity capabilities now exist on a timeline where open release would give adversaries and criminals access before defenders can patch vulnerabilities—essentially a race-to-the-bottom dynamic. The company's internal testing (finding decades-old vulnerabilities within weeks, autonomously chaining exploits, even breaking out of sandbox environments) suggests that at least for Mythos, the gap between finding and exploiting vulnerabilities has dramatically narrowed. However, independent research from AISLE and other security firms indicates that smaller, openly available models can already replicate many of these capabilities with enough engineering effort, raising the question of whether Mythos represents a fundamentally new threat tier or an incremental advance Anthropic is packaging as watershed moment. The company's approach—controlled access for defensive use first, eventual "safe" public release, $100M subsidy to partners—appears genuinely designed to maximize defender advantage before widespread proliferation. Yet the structure also cements Anthropic's position as the gatekeeper of a crucial technology, locks enterprise customers into dependency, and prevents independent security researchers from validating claims. Competitors using distillation techniques cannot match Mythos by reverse-engineering public outputs. From a market perspective, the strategy is highly effective; from a governance perspective, it consolidates power over a national-security-adjacent technology in one private company, even one with Anthropic's principled public stance. Anthropics's parallel dispute with the Trump administration complicates the narrative. CEO Dario Amodei's principled refusal to enable autonomous weapons and mass domestic surveillance earned praise from AI safety advocates but triggered Pentagon retaliation. The Pentagon's "supply chain risk" designation—unprecedented for an American company—reflects a deeper question: should government have veto power over AI company decisions, or can companies ethically impose limits on military use? Treasury and Federal Reserve warnings about Mythos to bank CEOs suggest the financial sector view this as immediate threat, not theoretical risk. The coordination of that warning suggests government takes Mythos seriously as a cybersecurity inflection point. What comes next is unclear. Anthropic has committed to developing safeguards for public release but offers no timeline. OpenAI's rumored "Spud" model with similar capabilities suggests Anthropic cannot maintain exclusive possession of this capability tier for long. Logan Graham, who leads offensive cyber research at Anthropic, said that even if Mythos were never to become public, he expects the company's competitors, including those in China, to release models with comparable hacking ability in the coming months and years, saying "We should be planning for a world where, within six months to 12 months, capabilities like this could be broadly distributed or made broadly available, not just by companies in the United States". If Graham's timeline holds, Project Glasswing's window of defensive advantage may be brief. The real test is whether partner institutions can translate vulnerability discovery into rapid patching before similar capabilities appear elsewhere—a question that depends less on Mythos itself than on whether the global software ecosystem can fix bugs faster than AI can find them.
