Anthropic announced Claude Mythos and project Glasswing on April 7th.
Claude Mythos is a new large language model which is reported to be exceptionally good at coding, reasoning and cyber security. The model is not public at this time so it's difficult to say how much is simply marketing buzz, but Anthropic claim the model has "already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser"1.
Project Glasswing is the consortium of some of the largest tech businesses in the world which have started using the Claude Mythos model to assist with defensive software engineering – finding vulnerabilities and weaknesses in code before they go live.
That's all nice, but let's get right to the big question:
There is a lot of noise and worry around the rise of highly capable automated tooling and Mythos specifically in the last week, and all the actions we 'need' to take as organisations to defend against them.
To jump straight to the point: There is no immediate cause for worry above what was already warranted.
That's not to say there's nothing to do, but that the actions that organisations should already be taking should simply be redoubled.
Over the last few years, the raise of AI assistance and automation has made models like Claude Mythos an inevitability, but there is often a tendency to make everything sound scary in cyber security. This is just the latest wake-up call in the ever lowering bar to entry for malicious actors conducting offensive activities. Ultimately, the time-to-exploit unpatched or previously unknown vulnerabilities and weaknesses is going to continue reducing, as large language models are further trained and optimised on datasets that include large numbers of codebases as well as offensive (and defensive) cyber security tooling.
There is nothing at this time to indicate Claude Mythos or any other model is able to achieve things a skilled human, given enough time and determination, couldn't also achieve.
There are certainly many actions that could be taken away as part of a broader approach to enhancing resilience against increasing numbers of cyber attacks. However, much like the summary above, let's cut right to the chase on the most important factor here:
The endless game of cat and mouse between attackers and defenders continues; models like Mythos show attackers are getting faster, so defenders need to be too.
The above is, of course, a simplification. To achieve it an organisation must first have a very good grasp of what's on the external perimeter, often easier said than done, especially for larger organisations. This then needs to be coupled with continuous monitoring and scanning to detect when something is out-of-date, and finally it needs a way to rapidly patch (or where no patch is available, assess and remove or isolate) any vulnerable exposed services.
Once the external perimeter is covered, attention should turn to insider threats - the disgruntled employee or a malicious actor who has somehow gained access to the internal side of an organisation. The same applies here: identify assets, monitor them, and patch them at speed.
Another knock-on effect is that our traditional approach to viewing threat actor capabilities in traditional threat models should be re-evaluated to include higher capabilities at lower levels than before. A script kiddie is in the near future going to be more capable than a script kiddie a few years ago who may have at best been able to muster up running some out-of-the-box Metasploit modules and vulnerability scanners.
A major concern brought to the front as part of these conversations is zero days - new, previously unknown exploits with no known fix, which Claude Mythos is reportedly very good at finding. Zero days have always been a problem, although there is likely to be an uptick in the number of them as it becomes easier to search for vulnerabilities.
There is nothing new to advise here; being informed and ready to act quickly when a zero day exploit affecting services used by the organisation is in the wild, then taking measures such as disabling the service or finding other ways to mitigate are the order of the day.
However, this is why organisations should never rely on a single line of defence. If a zero day can't be stopped from being exploited, then broader cyber resilience should be considered. This includes other defences that should kick in the moment there is any abnormal activity, escalation, traversal, conditional access violations, even falling back to segmentation to limit blast radius and the myriad other options we have within our cyber security arsenal.
The average business which may consider itself at risk to advanced offensive tools is covered above, but many business, and many of our clients, actively develop their own products.
For organisations developing software and products, again focusing on a single takeaway, the most important consideration as a result of recent news is:
Start leverage those same detection techniques as the attackers in the development pipeline.
Everyone wants to 'find it before it goes live' when it comes to vulnerabilities. Detecting every possible issue is much easier said than done and warrants its own article on modern DevSecOps pipelines with mature dynamic and static application security tests at multiple stages of the development cycle, starting as far 'left' as possible (before the developer even commits the code for review) and ensuring a combination of human and machine detections are in place.
Conducting automated security vulnerability analysis using AI assistants at multiple stages within development needs to become as second nature as all the linting, regression, UAT and other functional testing that already take place in various environments.
Project Glasswing, which was announced alongside Claude Mythos is a consortium of some of the largest technology companies which have been using Claude Mythos to help improve their secure development practises, is an excellent example of using tools defensively before any product is live for attackers to even try and poke holes in.
In the week following, OpenAI also announced their Trusted Access for Cyber (TAC), an initial version of which is released as part of GPT-5.4-cyber. According to OpenAI, the model "enables new capabilities for advanced defensive workflows"2 and is tuned for defensive cyber security use-cases.
Let's not pretend otherwise: there is no magic bullet new security strategy to adopt in this article, just like all the others pontificating about Mythos and its impact. The key takeaway is to detect and patch faster as attackers are faster, and ensure we as organisations are focusing on our cyber resilience. We could expand on that a bit to come up with the following takeaways:
If you would like to discuss this further, please contact us to speak to a member of the team.
1 Project Glasswing: Securing critical software for the AI era \ Anthropic
2 https://openai.com/index/scaling-trusted-access-for-cyber-defense/