Identifying vulnerabilities and exploits with Claude Mythos Preview
Over the previous few weeks, we’ve used Claude Mythos Preview to determine hundreds of zero-day vulnerabilities (that’s, flaws that had been beforehand unknown to the software’s builders), a lot of them critical, in each main working system and each main net browser, together with a variety of different necessary items of software.
In a publish on our Frontier Red Team blog, we offer technical particulars for a subset of those vulnerabilities which have already been patched and, in some instances, the ways in which Mythos Preview discovered to take advantage of them. It was capable of determine practically all of those vulnerabilities—and develop many associated exploits—solely autonomously, with none human steering. The following are three examples:
- Mythos Preview discovered a 27-year-old vulnerability in OpenBSD—which has a fame as certainly one of the most security-hardened working methods in the world and is used to run firewalls and different critical infrastructure. The vulnerability allowed an attacker to remotely crash any machine operating the working system simply by connecting to it;
- It additionally found a 16-year-old vulnerability in FFmpeg—which is utilized by innumerable items of software to encode and decode video—in a line of code that automated testing instruments had hit 5 million occasions with out ever catching the downside;
- The mannequin autonomously discovered and chained collectively a number of vulnerabilities in the Linux kernel—the software that runs most of the world’s servers—to permit an attacker to escalate from peculiar consumer entry to finish management of the machine.
We have reported the above vulnerabilities to the maintainers of the related software, and so they have all now been patched. For many different vulnerabilities, we’re offering a cryptographic hash of the particulars right this moment (see the Red Team weblog), and we’ll reveal the specifics after a repair is in place.
Evaluation benchmarks comparable to CyberGym reinforce the substantial distinction between Mythos Preview and our next-best mannequin, Claude Opus 4.6:
Cybersecurity Vulnerability Reproduction
In addition to our personal work, a lot of our companions have already been utilizing Claude Mythos Preview for a number of weeks. This is what they’ve discovered:
The highly effective cyber capabilities of Claude Mythos Preview are a results of its robust agentic coding and reasoning abilities. For instance, as proven in the analysis outcomes under, the mannequin has the highest scores of any mannequin but developed on quite a lot of software coding duties.
More data on the mannequin’s capabilities, its security properties, and its common traits will be present in the Claude Mythos Preview system card.
We don’t plan to make Claude Mythos Preview usually out there, however our eventual objective is to allow our customers to soundly deploy Mythos-class fashions at scale—for cybersecurity functions, but additionally for the myriad different advantages that such extremely succesful fashions will deliver. To achieve this, we have to make progress in growing cybersecurity (and different) safeguards that detect and block the mannequin’s most harmful outputs. We plan to launch new safeguards with an upcoming Claude Opus mannequin, permitting us to enhance and refine them with a mannequin that doesn’t pose the identical degree of danger as Mythos Preview3.
Plans for Project Glasswing
Today’s announcement is the starting of a longer-term effort. To achieve success, it’ll require broad involvement from throughout the know-how {industry} and past.
Project Glasswing companions will obtain entry to Claude Mythos Preview to search out and repair vulnerabilities or weaknesses of their foundational methods—methods that characterize a really massive portion of the world’s shared cyberattack floor. We anticipate this work will give attention to duties like native vulnerability detection, black field testing of binaries, securing endpoints, and penetration testing of methods.
Anthropic’s dedication of $100M in mannequin utilization credit to Project Glasswing and extra contributors will cowl substantial utilization all through this analysis preview. Afterward, Claude Mythos Preview will likely be out there to contributors at $25/$125 per million enter/output tokens (contributors can entry the mannequin on the Claude API, Amazon Bedrock, Google Cloud’s Vertex AI, and Microsoft Foundry).
In addition to our dedication of mannequin utilization credit, we’ve donated $2.5M to Alpha-Omega and OpenSSF by means of the Linux Foundation, and $1.5M to the Apache Software Foundation to allow the maintainers of open-source software to answer this altering panorama (maintainers focused on entry can apply by means of the Claude for Open Source program).
We intend for this work to develop in scope and proceed for many months, and we’ll share as a lot as we are able to in order that different organizations can apply the classes to their very own safety. Partners will, to the extent they’re ready, share data and finest practices with one another; inside 90 days, Anthropic will report publicly on what we’ve realized, in addition to the vulnerabilities mounted and enhancements made that may be disclosed. We will even collaborate with main safety organizations to provide a set of sensible suggestions for how safety practices ought to evolve in the AI era. This will probably embody:
- Vulnerability disclosure processes;
- Software replace processes;
- Open-source and supply-chain safety;
- Software improvement lifecycle and secure-by-design practices;
- Standards for regulated industries;
- Triage scaling and automation; and
- Patching automation.
Anthropic has additionally been in ongoing discussions with US authorities officers about Claude Mythos Preview and its offensive and defensive cyber capabilities. As we famous above, securing critical infrastructure is a prime nationwide safety precedence for democratic international locations—the emergence of those cyber capabilities is one more reason why the US and its allies should keep a decisive lead in AI know-how. Governments have an important position to play in serving to keep that lead, and in each assessing and mitigating the nationwide safety dangers related to AI fashions. We are able to work with native, state, and federal representatives to help in these duties.
We are hopeful that Project Glasswing can seed a bigger effort throughout {industry} and the public sector, with all events serving to to handle the greatest questions round the affect of highly effective fashions on safety. We invite different AI {industry} members to hitch us in serving to to set the requirements for the {industry}. In the medium time period, an impartial, third-party physique—one that may deliver collectively private- and public-sector organizations—is perhaps the very best dwelling for continued work on these large-scale cybersecurity tasks.
