After Anthropic's "Mythos" AI shocked the security world by unearthing decades-old, root-access flaws buried in the software running your bank, your browser, and your phone, Microsoft is racing to build a cheaper challenger called Project Perception — and the fight between them could decide who controls the future of digital defense.
Seattle / San Francisco
Somewhere inside FreeBSD's network file-sharing code, a flaw had been sitting quietly for seventeen years. It let anyone on the internet walk in through an unauthenticated back door and seize full root access to an affected machine — the digital equivalent of a master key to a bank vault, lost in a hedge for nearly two decades, undiscovered by generations of human security researchers, government auditors, and open-source volunteers who had all, at various points, looked directly at that code.
Then, earlier this year, an AI model found it in a matter of hours.
That model was Claude Mythos, a specialized cybersecurity system built by Anthropic and unveiled this spring, and its debut sent a very specific kind of shiver through the global security industry — not because anyone doubted AI would eventually get good at finding software bugs, but because of how quickly, and how completely, the theoretical became real. Now, months later, Microsoft is preparing to answer with its own rival system, reportedly called Project Perception, aiming to do essentially the same job at a fraction of the price — and setting up one of the most consequential corporate rivalries of the AI era, one where the prize isn't market share so much as the integrity of the software the entire world already runs on.
The Model That Found What Nobody Else Could
To understand why Microsoft is scrambling, you have to understand what Mythos actually did.
Announced by Anthropic on April 7, Claude Mythos Preview is what the company describes as a frontier-level cybersecurity model — not a general chatbot with security knowledge bolted on, but a system purpose-built for one narrow, extraordinarily high-stakes task: reading a piece of software's source code, figuring out where its weaknesses are likely to be hiding, and then proving those weaknesses are real by actually building a working exploit for them.
The workflow, as described by security researchers who've studied the model, looks unnervingly close to what an elite human penetration tester does, just compressed into a fraction of the time. Mythos reads a target codebase, ranks individual files by how likely they are to contain an exploitable flaw, generates specific hypotheses about where and how a vulnerability might exist, then actually runs the software to test whether each hypothesis holds up. When it finds a real flaw, it doesn't just flag it — it produces a full bug report, complete with a working proof-of-concept exploit and the exact steps needed to reproduce it.
In a controlled evaluation period lasting roughly seven weeks, according to reporting that has circulated widely in security circles, the model surfaced more than 2,000 previously unknown, high-severity software vulnerabilities. The list of what it found reads like a highlight reel of software's oldest, best-hidden sins: alongside that 17-year-old FreeBSD flaw, researchers point to a 27-year-old crash vulnerability in OpenBSD — an operating system whose entire cultural identity is built around obsessive, paranoid security auditing — that had somehow survived every single security review conducted on it since 1999. Mythos also reportedly surfaced a 16-year-old flaw buried in FFmpeg's H.264 video codec decoder, a piece of software so widely embedded in other programs that it's likely running, invisibly, on hardware you own right now.
Anthropic, to its credit, did not simply release this thing onto the open internet. The company has been unusually candid that a tool this good at finding exploitable bugs is also, definitionally, a tool that's extremely good at finding ways to break into things — and that putting it in the wrong hands could do serious damage before defenders even knew what hit them. Instead, Anthropic built what it calls Project Glasswing, named after a species of butterfly whose wings are naturally transparent, hiding in plain sight — a coalition designed to make sure the benefits of the discovery flow to defenders faster than they can be turned against them.
Project Glasswing's launch partner list reads like a who's-who of the companies that keep the internet's infrastructure standing: Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks all signed on at launch, with Anthropic later extending access to more than 150 additional partner organizations across over 15 countries — essentially any company or institution that builds or maintains software infrastructure critical enough that a hidden flaw in it could ripple out into a genuine crisis. Anthropic has backed the effort with real money, too: roughly $100 million in usage credits for participating organizations, plus a further $4 million in direct donations to open-source security groups, including gifts to the Open Source Security Foundation and the Apache Software Foundation specifically earmarked to help unpaid, overstretched open-source maintainers cope with a security landscape that just got a lot more demanding.
Access to Mythos itself isn't free, and it isn't cheap. Through Project Glasswing, participating organizations can reach the model via the Claude API, Amazon's Bedrock platform, Google Cloud's Vertex AI, or Microsoft's own Foundry service, at a price of roughly $25 per million input tokens and $125 per million output tokens — pricing that industry estimates place at roughly double the cost of Anthropic's flagship Opus model, and well above even the priciest tiers of rival systems from OpenAI.
Enter Microsoft — With a Cheaper Pitch
That price tag is precisely the opening Microsoft appears to be aiming for.
According to reporting from The Information and multiple technology outlets, Microsoft has been quietly developing its own answer, internally known as Project Perception, expected to launch as soon as this month. Functionally, the pitch is similar to Mythos: a system deployed inside an organization's own IT environment to scan for vulnerabilities and suggest fixes before attackers find them first. The headline difference Microsoft is selling isn't raw capability — it's cost.
Rather than relying on a single, expensive frontier model for every task, Project Perception reportedly uses a routing system that matches each individual security job to whichever AI model is best suited and cheapest for that specific task, drawing on a mix of models from Anthropic, OpenAI, and Microsoft's own AI stack. It's a strategy that echoes a broader trend across the tech industry, where companies increasingly blend expensive frontier models with cheaper, often open-source alternatives — including a growing number of lower-cost Chinese models — to keep the overall bill for running AI services under control. The project is reportedly being spearheaded by Hayete Gallot, who took over as Microsoft's head of security in February and has been restructuring the company's security business around AI-driven products, pulling investment away from older, non-AI tools in the process.
The competitive subtext here is impossible to miss, and Microsoft's own leadership hasn't been shy about drawing it out into the open. The company has been simultaneously trying to woo enterprise customers away from the very AI labs — OpenAI and Anthropic — that it has invested billions of dollars into and partnered with for years. Microsoft CEO Satya Nadella made headlines earlier this month with pointed comments warning that rival AI labs were, in his telling, using customer data to improve their own models in ways that raised governance concerns — while pointedly suggesting Microsoft had stronger internal controls in place to prevent similar practices. Microsoft's AI lead, Mustafa Suleyman, has separately said he wants the company to be launching genuinely frontier-level AI models of its own within the next year or so, rather than continuing to build products on top of models licensed from outside partners.
Layered on top of the commercial rivalry is a geopolitical one. Anthropic's own Fable 5 model — trained using techniques similar to Mythos but with additional safety guardrails layered on — was reportedly restricted from export outside the United States amid national-security-style scrutiny, forcing Anthropic to temporarily pull back access while it worked through a government review process to demonstrate the model was safe to release more broadly. Microsoft, notably, imposed its own restrictions on Fable access for its Azure cloud customers around the same time. Yet demand for Mythos-style capability hasn't cooled at all — multiple U.S. government agencies reportedly continue using it for cyber operations even as officials in Washington have separately floated designating Anthropic something close to a "supply risk," while governments and financial institutions across Europe have simultaneously been trying to secure their own access to the same technology. Microsoft, as a company less entangled in that specific export-control fight, may find itself with a real advantage in bringing a Mythos-style tool to European and other international customers faster and with fewer complications than Anthropic currently faces.
Why This Terrifies — And Reassures — Security Experts at the Same Time
Ask cybersecurity professionals how they feel about all this and you'll get a genuinely split answer, and it's worth sitting with why.
On one hand, there is real, grounded alarm. These models represent, in the words of researchers who've reviewed Mythos's capabilities, a moment where AI systems have crossed a threshold into being able to outperform all but the most elite human security researchers at the specific task of finding and weaponizing software flaws. Every capability that helps a defender patch a hidden vulnerability before anyone exploits it is, by the exact same logic, a capability that could help an attacker find that same vulnerability first, if the tool ever escapes the guardrails its creators have built around it — or if a less careful company builds a similar tool without equivalent safeguards. Commentators have described the reaction inside parts of the cybersecurity industry as bordering on outright hysteria at points earlier this year, even as more sober analysts have argued the underlying threat — AI-accelerated vulnerability discovery — was already arriving regardless of what any single company chose to do, making the real question not whether this technology would exist, but who would control it and how responsibly.
On the other hand, there's genuine, if cautious, optimism. For decades, cybersecurity has been fundamentally asymmetric in favor of attackers: a company has to defend every single door in its digital house, while an attacker only has to find one unlocked window. A tool that can comb through millions of lines of code and systematically flag the equivalent of unlocked windows, at a fraction of the time and cost a human audit would take, has the potential to flip that asymmetry, at least partially, back in favor of the defenders — provided it reaches them faster and more cheaply than it reaches the people trying to break in.
That's precisely the gap Microsoft's Project Perception is aiming to widen. A second credible player in this market, competing hard on price, doesn't just create a horse race between two Big Tech giants — it plausibly means AI-assisted vulnerability detection becomes financially reachable for the mid-sized banks, hospital networks, and infrastructure operators who could never have justified Mythos's premium pricing, but who are exactly the kind of organizations that have historically been softest targets for cybercriminals precisely because they couldn't afford serious security tooling.
What This Means for You
If you're not a network administrator or a chief information security officer, this might feel like an argument between billionaires and their AI labs that has nothing to do with your Tuesday. It has more to do with it than you'd think.
Nearly every piece of consumer technology you touch — your phone's operating system, the browser you're reading this in, the video streaming app you used last night, the banking app that holds your money — is built on layers of shared, often decades-old open-source code, the exact kind of code these models are being turned loose on. A flaw like the 17-year-old FreeBSD bug or the 27-year-old OpenBSD crash vulnerability doesn't sit in some abstract corporate server; in various forms, code like it quietly underpins huge swaths of the internet's actual infrastructure, including systems that route your data, verify your identity, and protect your payment information. When companies like Anthropic and Microsoft race to find and patch flaws like these faster than criminals can find and exploit them, the practical result — assuming the responsible-disclosure model holds — is fewer of the mass data breaches, ransomware outbreaks, and identity-theft waves that have made headlines with grim regularity over the past several years.
There's a flip side worth understanding, too. As these tools become cheaper and more widely available — which is precisely what a price war between Microsoft and Anthropic would produce — the same underlying technology inevitably becomes more accessible to less scrupulous actors as well, whether that's less careful competitors building similar tools without Anthropic's safeguards, or state-linked hacking groups adapting the same techniques for offense rather than defense. Security researchers are already treating "AI-discovered zero-day vulnerabilities" as a permanent new category of internet risk, not a temporary novelty, and consumers should expect security patches for the software they use to arrive both more frequently and, at least at first, in larger and more urgent batches than the usual monthly rhythm of updates most people are used to ignoring.
The Road Ahead
Project Perception has not yet formally launched, and Microsoft has not detailed exact pricing, though multiple reports suggest the company may impose access restrictions similar to those Anthropic has placed around Mythos, treating it as a carefully gated tool for vetted organizations rather than an open commercial product. Anthropic, for its part, shows no sign of slowing down Project Glasswing's expansion, having already grown the coalition from a dozen launch partners to more than 150 organizations across 15-plus countries within months of the program's debut.
What's clear is that the era of purely human-paced software security auditing is ending, quietly and much faster than most people outside the industry have noticed. The companies that build the phones, browsers, banking apps, and cloud platforms billions of people rely on every day are now in an active arms race to find their own flaws before someone else does it for them, with malicious intent. Whether that race is won by Anthropic's Mythos, Microsoft's Project Perception, or whatever OpenAI, Google, or a scrappier challenger builds next, the outcome will shape how safe your data actually is — not someday, but this year.
