A number of interesting data points have come out over the last few months about the latest foundation models and their bug/vulnerability finding capabilities, Agent Security, time-to-exploit for Zero-Days, and just this last few weeks the OpenAI/Hugging-Face Incident.
Going to try to pull some of these together, ask a few provocative comments, and figure out the implications.
Tl;dr - Foundation Models along with Agents used for attacks will fundamentally challenge our current assumptions about internet security and threat modeling. Robert Graham is one of the folks I think who has most clearly said - “By any rational measure, the Internet is indeed secure enough.” Foundation Models paired with Agents as attackers might be about to fundamentally change that and upset the apple cart.
Let’s look at some of the underlying shifts that are happening right now. Historically, the data has shown that most bugs and vulnerabilities in the real world simply never get exploited. In fact, many organizations are so confident in this empirical trend that they’re putting their money where their mouth is, relying on warranties and insurance like those from Root Evidence. Because we can’t possibly patch every single CVE without spending all of our time chasing our tails, we’ve leaned on the assumption that we only need to worry about the ones that actually present an active threat.
But that assumption is starting to crack. While there is a lot of hype around AI tools finding bugs, today they mostly rely on having direct access to source code. It’s one thing to have a model inside your environment scanning your own codebase—where I’d argue the bigger risk is still vulnerabilities in the open-source dependencies you use—but it’s quite another for an external attacker to gain that same visibility. That said, even without looking at source code, AI agents are becoming incredibly efficient at exploiting the vulnerabilities that are already known.
This shift means our threat modeling has suddenly become much more complicated. Most people don’t need to turn on "Lockdown Mode" on their iPhones because it’s a specific defense against high-tier nation-state actors like the NSO Group, and it comes with significant usability trade-offs. However, if automated agents can start attacking any vulnerability they find, the luxury of ignoring "unexploited" bugs disappears. We’re moving toward a world where the sheer scale of automated attacks could make the current state of the internet—where we all just live with thousands of bugs—completely untenable.
So, we have to ask ourselves which world we’re actually living in now. Are we still in a place where we can rely on historical data to prioritize only a handful of vulnerabilities, or are we entering a new reality where any and every vulnerability is a target waiting to be hit?
