Anthropic announced recently that their new Mythos model is too dangerous to release to the public. It can find vulnerabilities and exploit them at a rate never before seen. And while many of us are taking this news with a grain of salt since all the AI companies are in a hype race as much as they’re in a technology race, the reality is that AI is rapidly accelerating us towards a future where any target can be exploited in moments.
If you work in cybersecurity, this isn’t a lightbulb moment. AI acceleration of security research, proof, and execution has been a focus for years. Although many of these articles and papers are hard to find, pre ChatGPT era there were already experiments being run at major institutions designed around automated systems, both offensive and defensive. The release of LLMs and the accompanying surge of funding and attention has only accelerated the pace.
Even today, modern LLMs are incredibly effective at developing harnesses and performing rounds of research, proof of concept, and execution. You don’t have to look farther than some of the DARPA AI x Cyber challenge participants. The ARTIPHISHELL tool from Team Shellphish is an excellent example of an AI accelerated research, proof, and suggested remediation tool chain.
If you need to kick the tires on this, ask your AI tooling to perform security hardening on a sample project. Even if you already have it in your skills and prompts, you’ll be surprised at what a dedicated discovery round will find, even using the same model. And it isn’t even the fact that it will perform the additional sweep better and faster, it is that it can perform hundreds of code evaluations before you’ve even finished your coffee.
We are rapidly entering an epoch where software vulnerability is simply an accepted state of operating because there won’t be a way out of it. Some are thinking this is an apocalyptic state, where technology will simply collapse in on itself. For certain, that’s one path. I think there’s another that’s far more interesting and paints a future of a different, more secure overall environment, and it’s based on technologies that we’re already running now.
Whether you call it true cloud native, ephemeral computing, or ubiquitous compute, the response to a software environment filled with risk isn’t to double down on existing patterns that are the Monolith. A traditional, persistent, and stateful enterprise environment. Rather, we should lean into the idea of trusted baselines, anomaly detection, micro-segmentation, and ephemeral constructs. The system, out of compliance, should be able to heal itself by immediately and automatically destroying and reconstituting a clean, compliant instance. In research circles, there’s a lot of discussion about the AI and agentic layer needed to operationalize the monitoring and corrective actions. From an engineering perspective, there are things you should be doing with your teams today to prepare them for that future.
Although “serverless” is pretty much the norm these days, what counts as serverless or cloud native vastly depends on who you’re talking with and the context in which they are deploying. State persistence is the ally of the attacker, so containers won’t really cut it. The ability to load execution instructions onto a random compute target, perform the transform, and bail out will be a critical advantage. And some constructs do this today. All the major CSPs have some version of it: Google Cloud Run Functions, Azure Functions, AWS Lambdas, and there are even some open source projects.
Do these stateless compute platforms cover all the use cases? Not at all. This is the gap, then; finding ways to develop truly ephemeral compute that is constituted at the time of need for the business function, and then dissembled until the next time it is needed. The ability to rapidly constitute, dissemble, and reconstitute compute is the only metric that matters.
Is your enterprise ready for truly stateless operation? What do you need to be thinking about to prepare for that future?