How AI is getting higher at discovering safety holes : NPR

[ad_1] <img src="https://npr.brightspotcdn.com/dims3/default/strip/false/crop/5000x3336+0+0/resize/1100/quality/50/format/jpeg/?url=http%3A%2F%2Fnpr-brightspot.s3.amazonaws.com%2F08%2F6a%2F8c56ef0d4550b4a474c2abe4f42a%2Fap26057633599982.jpg" data-template="https://npr.brightspotcdn.com/dims3/default/strip/false/crop/5000x3336+0+0/resize/{width}/quality/{quality}/format/{format}/?url=http%3A%2F%2Fnpr-brightspot.s3.amazonaws.com%2F08%2F6a%2F8c56ef0d4550b4a474c2abe4f42a%2Fap26057633599982.jpg" class="img" alt="AI lab Anthropic announced this week that it had developed a powerful new model the company believes could “reshape cybersecurity.” It said that its latest model, Mythos Preview, was able to find "high-severity vulnerabilities, including some in every major operating system and web browser."" fetchpriority="high"/> AI lab Anthropic introduced this week that it had developed a robust new mannequin the corporate believes might "reshape cybersecurity." It stated that its newest mannequin, Mythos Preview, was capable of finding "high-severity vulnerabilities, together with some in each main working system and net browser." Patrick Sison/AP cover caption

toggle caption

Patrick Sison/AP Prior to now few months, AI fashions have gone from producing hallucinations to changing into efficient at discovering safety flaws in software program, in response to builders who preserve broadly used cyber infrastructure. These items of software program, amongst different issues, energy working programs and switch information for issues related to the web.

Whereas these new capabilities will help builders make software program safer, they can be weaponized by hackers and nation states to steal data and cash or disrupt essential providers.

The newest growth of AI's cyber functionality got here on Tuesday, when AI lab Anthropic introduced it had developed a robust new mannequin the corporate believes might "reshape cybersecurity." It stated that its newest mannequin, Mythos Preview, was capable of finding "high-severity vulnerabilities, together with some in each main working system and net browser." Not solely that, the mannequin was higher at developing with methods to use the vulnerabilities it discovered, which implies malicious actors can extra successfully obtain their objectives.

For now, the corporate is limiting the entry to the mannequin to round 50 choose firms and organizations "in an effort to safe the world's most important software program." They're calling the collaboration Venture Glasswing, naming it after a butterfly species with clear wings.

Anthropic says the danger for misuse is so excessive that it has no plans to launch this specific mannequin to most people, in response to the announcement, however it'll launch different associated fashions. "Our eventual aim is to allow our customers to securely deploy Mythos-class fashions at scale," the corporate wrote.

Nonetheless safety consultants say the potential dangers are for cybersecurity professionals and never on a regular basis folks. "I do not essentially assume that the common laptop person must be essentially nervous about this," stated Daniel Blackford, VP of Menace Analysis at Proofpoint, a cybersecurity agency. "They have to be far more nervous about not giving their password away as a result of that simply occurs like all day, day by day."

Mythos Preview may have higher capabilities to assist builders repair vulnerabilities than earlier fashions, stated Jim Zemlin, CEO of the Linux Basis, which hosts the open supply Linux kernel. A kernel is an interface that lets {hardware} discuss to software program, and the Linux kernel powers a few of the most generally used working programs together with Android, in addition to all the world's 500 strongest supercomputers. The muse is part of Venture Glasswing, and Zemlin stated a core group of people that work on the Linux kernel, referred to as maintainers, have began experimenting with the brand new mannequin to determine the simplest method to make use of it.

"These maintainers are already overworked earlier than AI," Zemlin stated. "This simply makes their lives so much higher."

Even earlier than Mythos Preview's restricted launch, the cybersecurity group has already been grappling with how succesful probably the most superior commercially accessible AI fashions have change into. Builders are racing to patch safety vulnerabilities that hackers flag with AI. Safety consultants are nervous about what may lie forward if such capacities proliferate with out guardrails.

Hackers are utilizing AI to assist them discover bugs. It wasn't working, till now Enchancment in AI fashions' capabilities grew to become noticeable early 2026, stated Daniel Stenberg. He's the lead software program developer behind cURL, a 30-year-old open-source information switch instrument that is broadly used for issues that hook up with the web, together with vehicles and medical gadgets. That change adopted the releases of recent cutting-edge fashions late 2025.

Like many within the software program group, Stenberg will get assist from safety researchers to maintain his software program useful and safe. So-called "white hat" hackers flag safety flaws to builders in non-public and in return, they generally get reward cash, referred to as "bug bounties," or bragging rights by getting the safety flaw named after themselves. Whereas all bugs have an effect on the performance of the software program, just some bugs current safety vulnerabilities.

However that course of began to interrupt down final 12 months. Stenberg's crew was inundated by bogus stories he was satisfied had been generated by AI. "We bought 185 stories all year long and fewer than 5% of them had been really safety associated issues in the long run," Stenberg stated.

Even because the report quantity doubled from 2024 to 2025, Stenberg ended up discovering and fixing fewer safety vulnerabilities. The slop stories bought so dangerous that Stenberg stopped paying out bug bounty rewards for his cURL instrument.

The hackers that file stories to him are normally nameless, stated Stenberg, they usually do not focus on whether or not the stories had been made with AI. However Stenberg stated AI has a definite writing model. "They are usually very elaborate and descriptive … You get a 400 line report [when] it is one thing {that a} human would take 50 traces to current."

HackerOne, a platform that Stenberg makes use of to handle safety reporting, surveyed hackers in the summertime of 2025 and located that almost 60% of respondents had been both utilizing AI, studying it, or studying to audit AI or machine studying programs.

"LLMs have now bypassed human functionality for bug discovering" Issues have modified dramatically this 12 months. The quantity of stories has been even larger than in 2025, however up to now Stenberg stated, most of them have uncovered professional points. "Nearly all of the dangerous [reports] at the moment are gone."

He estimates that about 1 in 10 of the stories are safety vulnerabilities, the remainder are largely actual bugs. Simply three months into 2026, the cURL crew Stenberg leads has discovered and stuck extra vulnerabilities than every of the earlier two years.

Apart from stories by exterior safety researchers, Stenberg additionally makes use of AI to seek out vulnerabilities himself. With one click on, AI has flagged over 100 bugs in his code that has gone by rounds of overview by people and conventional code analyzers "in nearly magical methods."

Stenberg's expertise is just not distinctive. Maintainers of the Linux kernel have seen an analogous change within the high quality of bug stories. Nicholas Carlini, an Anthropic analysis scientist, was capable of finding vulnerabilities within the Linux kernel utilizing an older Anthropic mannequin and a comparatively easy immediate. Carlini additionally used AI to seek out the primary essential vulnerability in one other 20-year-old open supply venture.

"LLMs have now bypassed human functionality for bug discovering," stated Alex Stamos, chief safety officer at Hall, an AI software program safety firm. The rise in high quality of the safety analysis stories adopted the discharge of Anthropic's mannequin Opus 4.5 in November, stated Stamos, who was beforehand the pinnacle of safety at Yahoo and Fb. Since a lot business software program has open-source elements, what occurs to open-source tasks has wider implications for the web, Stamos stated.

Can AI patch the bugs and safety vulnerabilities it finds? Stenberg appreciates that AI fashions at present accessible to everybody have change into extra useful find bugs, however he is additionally cautious of what future, extra highly effective fashions may convey for builders who preserve open-source software program. "It is an overload of all of the maintainers who're already usually overloaded and understaffed and underpaid and underfunded in some ways."

Stenberg is not a part of Venture Glasswing and says that a variety of essential tasks, "issues which might be really cornerstones of the Web," have been ignored.

Anthropic didn't reply to NPR's request for remark.

However based mostly on his expertise up to now, Stenberg says AI is just not nearly as good at fixing bugs and safety flaws as it's at discovering them.

A part of the reason being that neither the bugs nor their fixes are cut-and-dry. Like many different elements of software program growth, making the judgement calls takes extra time than really writing the code. "As soon as we've recognized the issue and agree that it is a drawback, then really fixing it's not very laborious and never very time consuming. It is extra the complete course of as much as that step that takes time and vitality," Stenberg stated.

Others assume in another way. An organization referred to as HackerOne is now growing an agentic AI product to extra autonomously discover— and mend— vulnerabilities.

How does AI issue into the tug of warfare between offense and protection in cybersecurity? "Discovering bugs is just not making the bugs exploitable," Stamos stated. "The primary a part of what we name the kill chain is discovering the issues. The subsequent step is… [actually] constructing the weapon. And the inspiration fashions won't do this for you."

Basis fashions are fashions made by probably the most superior AI labs like Anthropic, OpenAI and Google Deepmind. These labs have put guardrails in place to forestall their fashions from creating software program that can be utilized maliciously. These fashions are extremely proprietary and their internal workings aren't public.

However all bets are off as soon as so-called open-weight fashions which might be extra accessible to the general public catch as much as these from probably the most superior AI labs. Unhealthy actors could make a duplicate of them and take away the guardrails towards malware.

"Then we're in actual hassle since you would be capable to ask these fashions to not simply discover the bugs, however then to create exploit code" that may hack into programs, Stamos stated. Essentially the most superior open-weight fashions are lower than a 12 months behind probably the most superior closed-weight fashions.

"It additionally exhibits how extremely silly it's for the Pentagon to say that Anthropic is a danger to the complete United States," Stamos added. The Pentagon has labelled Anthropic a "provide chain danger" as a result of the corporate requested the federal government to not use its know-how for autonomous weapons and mass surveillance. The label would bar authorities businesses and contractors from working with Anthropic. Anthropic is at present disputing the label in court docket.

A lot of the open-weight fashions are made by firms based mostly in China, which the U.S. considers its foremost rival within the race in direction of AI dominance. By not releasing Mythos Preview publicly, Stamos stated, Anthropic is giving software program builders in addition to the U.S. time to shore up their defenses.